txt/2201.01661.txt

1 
 
  
Evaluation of Thermal  Imaging on Embedded GPU 
Platforms  for Application in Vehicular Assistance 
Systems    
 
Muhammad Ali Farooq, Waseem Shariff, Peter C orcoran, Fellow, IEEE   
Abstract—This study is focused on evaluating the real-time 
performance of thermal object detection for smart and safe 
vehicular systems by deploying the trained networks on GPU & 
single -board EDGE -GPU computing platforms for onboard 
automotive sensor suite testing. A novel large -scale thermal 
dataset comprising of > 35,000 distinct frames is acquired, 
processed, and open -sourced  in challenging weather and 
environmental scenarios . The dataset is a recorded  from lost -cost 
yet effective uncooled LWIR thermal camera , mounted stand -
alone and on an electric vehicle to minimize mechanical vibrations . 
State -of-the-art YOLO -V5 networks variants are trained using 
four different public datasets as well newly acquired local dataset 
for optimal generalization of DNN by employing SGD optimizer.  
The effectiveness of trained networks is validated on extensive test 
data u sing various quantitative metrics which include precision, 
recall curve, mean average precision, and frames per second. T he 
smaller network  variant  of YOLO is further optimized using 
TensorRT inference accelerator  to explicitly boost the frames per 
second rate. Optimized network engine increases the frames per 
second rate by 3.5 times when testing on low power edge devices  
thus achieving 11  fps on Nvidia Jetson Nano and 60 fps on Nvidia 
Xavier NX development boards . 
 
Index Terms — ADAS, Object detection, Thermal imaging, 
LWIR, CNN, Edge computing  
I. INTRODUCTION  
hermal imaging is the digital interpretation of the 
infrared radiations emitted from the object. Thermal 
imaging cameras with microbolometer focal plane 
arrays (FPA) is a type of uncooled detector that provides low -
cost solutions for acquiring thermal images in different weather 
and environmental conditions. These cameras when integrated 
with AI -based imaging pipelines can be used for various real-
world  applications.  In this work, the core focus is to design an 
intelligent thermal object detection -based video analysis system 
for automotive sensor suite application that should be effective 
in all light conditions thus enabling safe and more reli able road 
journeys. Unlike other video solutions such as visible imaging 
which mainly relies on reflected light thus having the greater 
chances of being blocked by visual impediments, thermal 
imaging does not require any external lighting conditions to 
capture quality images and it can see through visual obscurants 
 
October 25th, 2021 , “This research work is funded by  the ECSEL Joint 
Undertaking (JU) under grant agreement No 826131  (Heliaus project  
https://www.heliaus.eu/  ). The JU receives support from the European Union’s 
Horizon 2020 research and  innovation program and National funders  from 
France, Germany, Ireland (Enterprise Ireland , International Research Fund), 
and Italy ”.   such as dust, light fog, smoke, or other such occlusions. 
Moreover, the integration of AI-based thermal imaging systems 
can provide us with a multitude of advantages from better 
analytics with fe wer false alarms to increased coverage , provide  
redundancy  and, higher return on investment.  
In th is research work , we have focused on utilizing thermal 
data for designing efficient AI -based object detection and 
classification pipeline for Advance Driver -Assistance Systems.  
Such type of thermal imaging -based forward sensing (F -sense) 
system  is useful in providing enhance d safety and security 
feature s thus enabling the driver to better scrutinize the 
complete road -side environment.  For this purpose, we ha ve 
used a state-of-the-art (SoA)  end-to-end deep learning 
framework YOLO -V5 on thermal data. In the first phase , a 
novel thermal data set is acquired for training and validation 
purposes of different network variants of YOLO -V5. The data 
is captured using a  prototype low -cost uncooled LWIR thermal 
camera specifically designed under the ECSEL Heliaus  
research project [ 32]. The raw thermal data is processed using 
shutterless camera calibration, automatic gain control, bad -
pixel removal , and temporal denoising methods.  
Furthermore,  the trained network variants are deployed and 
tested on two state-of-the-art embedded GPU platforms,  which 
include NVIDIA  Jetson nano  [23] and Nvidia Jetson Xavier  NX 
[25]. Thus, studying the extensive real -time and on -board 
feasibility in terms of various quantitative metrics, inference 
time, FPS,  and hardware sensor temperatures.  
The core contributions  of the proposed research work are 
summarized below : 
• Preparation and annotation of a large o pen-access dataset of 
thermal images  captured in different weather and 
environmental conditions.  
• A detailed comparative evaluation of SoA object detection 
based on a modified  YOLO -V5 network , fine-tuned for 
thermal images  using this newly acquired dataset .  
• Model optimization using TensorRT inference accelerator 
to implement a fast  inference network on SoA embedded  
GPU boards  (Jetson, Xavier) with comparative evaluations . 
Muhammad Ali Farooq, Peter Corcoran,  and Waseem Shariff are with the 
National University of Ireland Galway, (NUIG), Coll ege of Science & 
Engineering  Galway, H91TK33, Ireland (e -mail: [email protected] ,    
[email protected] , [email protected] ).  
Thermal Dataset Link:  https://bit.ly/3tAkJ0J        
GitHub Link :  https://git hub.com/MAli -Farooq/Thermal -YOLO   T 2 
 
 • A determination of realistic frame rates that can be achieved 
for the rmal object detection on SoA embedded GPU 
platforms .  
II. BACK GROUND  
ADAS  (Advanced Driver Assistance Systems) are classified 
as AI -based intelligent systems integrated with core vehicular 
systems to assist the driver by providing a wide range of digital 
features for safe and reliable road journeys. Such type of system 
is designed by employing an array of electronic sensors and 
optical mixtures such as different types of cameras to identify 
surrounding impediments, driver faults, and reacts 
automatically.  
The second part of this section will mainly summarize the 
existing/ p ublished thermal datasets along with their respective 
attributes. These datasets can be effectively used for training 
and testing the machine learning algorithms for object detection 
in thermal spectrum for ADAS. The complete dataset details are 
provided i n Table I .  
TABLE I  
EXISTING THERMAL DATASETS  
A. Related Literature  
We can find numerous studies regarding the implementation 
of object detection algorithms using AI based conventional 
machine learning as well as deep learning algorithms. Such type 
of optical imaging -based systems system can be deployed and 
effectively used as forward sensing methods for ADAS. 
Advanced Driver -Assistance Systems (ADAS) is an active area 
of research that seeks to make road trips more safe and secure.  
Real time object detection pl ays a critical role to warn the driver 
thus allowing them to make timely decisions [ 8]. Ziyatdinov et 
al [8] proposed an automated system to detect road signs. This method uses the GTSRB dataset [ 20] to train on conventional 
machine learning algorithms whi ch include SVM, KNN , and 
Decision Trees classifier. The results proved that SVM and K – 
nearest neighbour ( k-NN) outperforms all other classifiers.  
Autonomous cars on the road require the abilities to 
consistently perceive and comprehend their surroundings  [9]. 
Oliver et al [ 9] presented a procedure to use Bernoulli particle 
filter, which is suitable for object identification because it can 
handle a wide range of sensor measurements as well as object 
appearance -disappearance . Gang Yan et al [ 10] proposed a 
novel method to use HOG to extract features  and AdaBoost and 
SVM classifiers to detect vehicles in real -time. The histogram 
of oriented gradients (HOG) is a feature extraction technique 
used for object detection in the domai n of computer vision  and 
machine learning . The study concluded that the AdaBoost 
classification technique performed slightly better than SVM 
since it uses the ensemble method. Authors in [ 11], proposed 
another approach to detect vehicles on road using HOG filters 
to again extract features from the frames and then classify them 
using support vector machines and decision tree classification 
algorithms. Furthermore, SVM achieved 93.75% accuracy, 
which outperformed decision tree accuracy on classifying the 
vehicles. These are some of the conventional machine learning 
object detection techniques used for driver assistance system till 
date.  The main drawback of traditional machine learning 
technique s is that  the features are extracted and predefined  prior 
to training and tes ting of the algorithms . When dealing with 
high-dimensional data, and with many classes conventional 
machine learning techniques are often ineffective [ 21].  
Deep learning approaches have emerged as more reliable and 
effective solutions than these classic approaches. There are 
many state -of-the-art pre -trained deep learning classifiers  and 
object detection models which can be retrain ed and rapidly 
deployed for  designing efficient forward sensing algorithms 
[22]. YOLO  (you only look once) object classifier  provides 
sufficient performance to operate at real -time speeds on 
conventional video data without compromising the overall 
detector precision [ 15]. Veta et al [12] presented a technique for 
detecting objects at a distance  by employing YOLO on low -
quality thermal images . Another research [ 13] focused on 
pedestrian detection  in thermal images using the histogram of 
gradient (HOG) and YOLO methods on FLIR [ 7] datas et and 
computed performance with a 70 % accuracy on test data using 
the intersection over union technique. Further, Rumi et al [ 14] 
proposed a real -time human detection technique using YOLO -
v3 on KAIST [ 5] thermal  dataset, achieving 95.5% average 
precision on test data. Authors in [ 16] proposed a human 
detection system using YOLO object detector. The  authors used  
their custom dataset recorded in different weather conditions 
using FLIR Therma -CAM P10 thermal camera.  
Focusing  on road-side objects, authors in [ 17] used YOLO -v2 
object detection model to enhance the recognition of tiny 
vehicle objects by combining low -level and high -level features  
of the image .  In [18], the authors proposed a deep learning -
based vehicle occupancy detection  system  in a parking lot using 
a thermal camera . In this study authors had establis hed that 
YOLO, Yolo -Conv, GoogleNet , and ResNet18 are 
computationally more efficient, take less processing time, and 
are suitable for real -time object detection.  In one of the  most 
recent studies [ 24], the efficacy of typical state -of-the-art object Datasets  Condition  
Day Night  Annotations  Objects  Total 
no. of 
frames  Image 
Resolution  
OSU 
Thermal 
[2] ✓ ✓ - Person, 
Cars, 
Poles  284 360 X 240  
CVC 
[19] ✓ ✓ - Person, 
Cars, 
Poles, 
Bicycle, 
Bus, 
Bikes  11K 640 X  480 
LITIV 
[3] - - - Person  6K 320 X 240  
TIV [ 4] - - - Person, 
Cars, 
Bicycle, 
Bat 63K 1024 X 
1024  
SCUT 
[6] - ✓ ✓ Person  211K  384 X 288  
FLIR 
[7] ✓ ✓ ✓ Person, 
Cars, 
Poles, 
Bicycle, 
Bus, 
Dog 14K 640 X 512  
KAIST 
[5] ✓ ✓ ✓ Person, 
Cars, 
Poles, 
Bicycle, 
Bus 95K 640 X 480  3 
 
 
detectors which includes  Faster R -CNN, SSD, Cascade R -
CNN, and YOLO -v3 was assessed by retrain ing them on a  
thermal dataset.  The results demonstrated that Yolo -v3 
outclassed other object SoA object detect ors.   
B. Object Detection on Edge Devices  
AI on edge devices benefit s us in various methods such that 
it speeds up decision -making, makes data processing more 
reliable, enhan ces user experience with hyper -personalization, 
and cuts down the costs . While machine learning models ha ve 
shown immense strength in diversified consumer electronic 
applications , the increased prevalence of AI  on edge  has 
contributed to the growth of spec ial-purpose embedded boards 
for various applications.  Such type of embedded boards  can 
achieve  AI inference at higher frames per second (fps) and low 
power usage . Some of these board includes Nvidia Jetson Nano, 
Nvidia Xavier, Google Coral, AWS DeepLens , and Intel AI-
Stick . Authors in [ 26-27] proposed a r aspberry pi-based edge 
computing system  to detect thermal objects. Sen Cao et al [ 28] 
developed a roadside object detector using KITTI dataset [ 29] 
by training an efficient and lightweight neural  network on 
Nvidia Jetson TX2 embedded GPU  [28].  
In another study [ 30] authors  proposed deep learning -based 
smart task scheduling for self -driving vehicles. This task 
management module was implemented  on multicore SoCs 
(Odroid Xu4 and Nvidia Jetson).   
The overall goal of this study is to analy se the real -time 
performance feasibility of Thermal -YOLO object detector by 
deploying on edge devices.  Different network variants of yolo -
v5 framework are trained and fine -tuned on thermal image data 
and implemented on the Nvidia Jetson Nano [ 23] and Nvidia 
Jetson Xavier NX [ 25]. These two platforms, although from the 
same manufacturer provide very differe nt levels of performance 
and may be regarded as close to current SoA in terms of 
performance for embedded neural inference algorithms.  
III. THERMAL DATA ACQUISITION AT SCALE FOR ADAS  
This section will mainly cover the thermal data collection 
process using the LWIR pr ototype thermal imaging camera. 
The overall data is consisting of more than 35K distinct thermal 
frames acquired in different weather and environmental 
conditions . The data collection process includes shutterless 
camera calibration and thermal data processing [36], using the 
Lynred Display Kit (LDK) [ 1], data collection methods , and 
overall dataset attributes with different weather and 
environmental conditions for comprehensive data formation.  
A. Prototype Thermal Camera  
For the proposed research  work we have utilized micro -
bolometer technology based uncooled thermal imaging camera 
developed under the HELIAUS project [ 32]. The main 
characteristic of this camera includes low -cost, lightweight and 
its sleek compact design thus allowing to easily integrate it with 
artificially intelligent imaging pipelines for building effective 
in-cabin driver -passenger monitoring and road mon itoring 
systems for ADAS. It enables us to capture high -quality thermal 
frames with low -power consumption thus proving the agility of 
configurations and data processing algorithms in real -time.   
Fig. 1 shows the prototype thermal camera. The technical 
specifications of the camera are as follows, the camera type is a QVGA long-wave infrared (LWIR) with a spectral range from 
8-14 µm and a camera resolution of 640 X 480 pixels. The focal 
length (f) of the camera is 7.5 mm, F -number is 1.2, the pixel 
pitch is 17 µm, and the power consumption is less than 950mW. 
The camera relates to a high-speed USB 3.0 (micro -USB)  port 
for the interface.  
 
 
 
 
 
 
Fig. 1 . LWIR thermal imaging module images from different 
view angles.  
 
The data is recorded using a specifically designed toolbox. The 
complete camera calibration process along with the data 
processing pipeline is explained in the next section.  
B. Shutterless Calibration and Real -time Data Processing  
This section will highlight the thermal camera calibration 
process for shutterless camera configuration along with real -
time data processing methods for converting the raw thermal 
data to refined output s. Shutterless technology allows uncooled 
IR engines and thermal imaging sensors to continuously operate 
without the need for a mechanical shutter for Non -Uniformity 
Correction (NUC) operations. Such type of technology 
provides proven and effective results  in poor visibility 
conditions ensuring good quality thermal frames in real -time 
testing situations. For this, we have used a low -cost blackbody 
source to provide three different constant reference temperature 
values referred to as T -ambient1 -BB1 (hot unif orm scene with 
temperature value of 40 degree centigrade), T -ambient1 -BB2 
(cold uniform scene with the temperature value of 20 degree 
centigrade), and T -ambient2 -BB1 (either hot or cold uniform 
scene but with different temperature value). The  imager  can 
store up to 50 snapshots and select the best uniform temperature 
scenes for calibration purposes. Fig. 2  shows the blackbody 
used for the thermal camera calibration .  
 
 
 
 
 
 
 
 
 
 
Fig. 2.  Thermal camera calibration a) blackbody source used 
for LWIR thermal camera calibration, b) uniform scene: 
temperature set to 40.01 degree centigrade.  
 
Once the uniform temperature images are recorded the images 
are loaded in camera SDK as shown in Fig. 3 to finally calibrate 
the shutterless camera stream. Fig. 4 shows the results before 
applying shutterless calibration and processed results using 
shutterless algorithms on thermal frame capture through the 
prototype thermal IR camera . (a) (b) 4 
 
   
Fig. 3. Prototype thermal camera SDK for loading constant 
reference temperatures values for shutterless camera 
calibration.  
 
 
 
 
 
 
 
 
 
 
 
Fig. 4.  Shutterless algorithm results on sample thermal frame 
captured from 640x480 LWIR thermal camera designed by 
Lynred France [ 1]. 
 
In the next phase, various real -time image processing -based 
correction methods are applied to convert the original thermal 
data to produce good -quality thermal frames. Fig. 5 shows the 
complete image processing pipeline.  
 
 
 
 
 
 
 
 
 
Fig. 5.  Thermal image correction pipeline  
 
As shown in Fig. 5 image processing pipeline consist of three 
different image correction methods which include gain 
correction, bad -pixel replacement, and temporal denoising. The 
further details of these methods are provided as follows.  
 
1) Gain Correction Automatic Gain Control (AGC)  
Thermal image detectors, based on flat panels, suff er 
from irregular gains due to the non -uniform amplifiers. 
To correct the irregular gains, a common yet effective 
technique referred to as automatic gain control is 
applied. It is usually based on the gain map. By 
averaging uniformly illuminated images wit hout any 
objects, the gain map is designed. By increasing the 
number of images for averaging provides a good gain -correction performance since the remained quantum 
noise in the gain map is reduced [ 1]. 
 
2) Bad Pixel Replacement (BPR)  
This is used to list bad pixels estimated at the calibration 
stage. It works by tracking potential new bad pixels by 
looking at pixel neighbourhood also known as the 
nearest neighbour method. Once it traces the bad pixels 
in the nearest neighbour it replaces them with good 
pixels. Fig. 6 demonstrates one  such example.   
 
 
 
 
 
 
 
 
 
 
Fig. 6.  Bad pixel replacement algorithm output on 
sample thermal frame, left side frame with some bad 
pixels and the right side is processed frame.  
 
3) Temporal Denoising (TD)  
The consistent reduction of image noise poses a 
frequently recurring problem in digitized thermal 
imaging systems and especially when it comes to un -
cooled thermal imagers [ 34]. To mitigate these 
limitations for better outputs different methods are used 
which include hardware as well  software -based image 
processing methods such as temporal and spatial 
denoising algorithms. The temporal denoising method is 
used to decrease the temporal noise between different 
frames of the video. In commercial solutions, it usually 
works by gathering m ultiple frames and averaging those 
frames to cancel out the random noise among the frames. 
In our data acquisition process, this method is used after 
applying the shutterless algorithm. Fig. 7  shows the 
sample thermal images in the form of outcomes after 
applying shutterless algorithms and all the image 
processing -based corrections methods as shown in Fig. 
5. 
 
 
 
 
 
 
 
 
 
 
 
Fig. 7. High -quality thermal frames after applying the 
shutterless calibration algorithm and image correction 
methods.  
Before  After   
Shutterl
ess 
thermal 
data 
Gain 
correction  
Bad-pixel 
replacement  
Temporal 
denoising  
Image processing 
pipeline  
Output  
Input  Output   
5 
 
 C. Data Collection Methods and Overall Dataset Attributes  
This section will highlight different data collection 
approaches adopted in this research work. The data is collected 
in two different approaches. In, the first approach (M -1) the data 
is gathered in an imm obile method by placing the camera at a 
fixed place. The camera is mounted on the tripod stand at a 
fixed height of nearly 30 inches such that the roadsides objects 
are covered in the video stream. The thermal video stream is 
recorded at 30 frames per seco nd (FPS). The data is recorded in 
different weather and environmental conditions. Fig. 8 shows 
the M -1 data acquisition setup. In the second method (M -2) the 
thermal imaging system is mounted over the car and data is 
acquired in the mobile method. The prime reason for collecting 
the data in two different methods is to bring variations and 
collect distinctive local data in different environmental and 
weather conditions. For this, a specialized waterproof camera 
housing case was designe d to hold the thermal camera in the 
correct position and angle to cover the entire roadside scene. 
The housing case is fixed on a suction -based tripod stand thus 
allowing us to easily fix and remove the complete structure 
from the car bonnet. The housing c ase also contains a visible 
camera to get initial visible images as reference data thus 
allowing us to adjust both the camera positions in proper  angle 
and field of view .  
 
 
Fig. 8.   Data Acquisition setup by placing the camera at a fixed 
place a) camera mounted on a tripod stand, b) complete daytime 
roadside view, c) video recording setup at 30fps, d) evening 
time alleyway view.  
 
Fig. 9 shows the camera housing case along with the initial 
data acquisition setup whereas  
 
                                
Fig. 9.  Data acquisition setup through car a) camera housing 
case holding thermal and visible camera, b) initial data 
acquisition testing phase.  
 
Fig. 10 shows the housing case fixed on tripod structure and 
complete M -2 acquisition setup mounted on the car. The overall 
dataset is acquired from Galway County Ireland. The data is 
collected in form of short video clips and more th an> 35,000 
unique thermal frames have been extracted from the recorde d 
video clips. The data is recorded in the daytime, evening time , 
and night -time which is distributed in the ratio of 44. 61%, 31.78%, and 23. 61% respectively  of overall data. The complete 
dataset attributes are summarized in Table II. The acquired data 
comprises  distinct stationary  classes , such as road signs and 
poles,  as well as moving object  classes  such as  pedestrians, cars, 
buses, bikes, and bicycles.  
 
 
Fig. 10.   Complete data acquisition setup mounted on the car a) 
camera housing fixed on a suction tripod stand, b) data 
acquisition kit from the front view, c) data acquisition kit from 
the side view.  
TABLE II  
NEW THERMAL DATASET  ATTRIBUTES  
Locally acquired dataset attributes   
Data 
collection 
method with 
frame 
properties  Total 
number 
of 
extracted 
frames  Processing 
Method  Environment  Time and 
weather 
conditions  
M-1 
Camera 
mounted at a 
fixed place  
 
96 dpi 
(horizontal 
and vertical 
resolution) 
with 
640x480 
image 
dimension  8,140  
 Shutterless, 
AGC, BPR, 
TD 
 Roadside  Daytime 
with cloudy 
weather  
680 Alleyway  Evening 
time cloudy 
weather  
4,790  Roadside  Night -time 
with light 
cloudy and 
windy 
weather  
M-2 
Camera 
mounted on 
the car  
(Driving 
condition)  
 
96 dpi 
(horizontal 
and vertical 
resolution) 
with 
640x480 
image 
dimension  
 9,600 Shutterless, 
AGC, BPR, 
TD 
 Industrial 
Park Daytime 
with clear 
weather 
and light 
foggy 
weather  
11,960  Downtown  Evening 
time with 
partially 
cloudy and 
windy 
weather  
4,600  Shutterless, 
AGC, BPR,  
& TD  Downtown  Night -time 
with clear 
weather 
conditions  
frames  Daytime: 
17,740 
(44.61%) Evening 
time: 
12,640 
(31.78%) Night -time: 
9,390 
(23.61%) Total: 
39,770   
Fig. 11 shows the six distinct sample of thermal frames captured 
in different environmental and weather conditions using M1 
and M2 methods.  These samples show different class object s 
such as buses, bicycles , poles, person,  and cars.  Most of these 
objects are found commonly on the roadside  thus providing the 
driver a comprehensive video analysis of car surroundings . 
(a) 
 (b) (c) (d) 
(a) (b) (a) (b) (c) Suction Tripod 
structure  
640x480 LWIR 
thermal camera  6 
 
  
 
 
Fig. 11.  Six different thermal samples acquired using LWIR 
640x480 prototype thermal camera showing various class 
objects.   
IV. PROPOSED METHODOLOGY  
This section will detail the proposed methodology and 
training outcomes from the various network variants tested in 
this study.  
A. Network Training and Learning Perspectives  
The overall training data comprises both locally and publicly 
available datasets. The complete training data is divided in the 
ratio of 50% - 50% where 50% of data is selected from locally 
acquired t hermal frames whereas the rest 50% of the training 
data leverages  from public datasets. Six distinct types of 
roadside objects for driving assistance are included in training 
and validations sets . These include b icycles, motor cycles , 
buses, cars, pedestrians or people, and static roadside  objects 
such as  poles  or road signs , as shown  in Fig 1 2. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Fig. 12.  Block diagram depicts the steps taken to evaluate the 
performance of Yolo v5 on local and public datasets.  
 Fig. 13 shows the class -wise data distribution. In the training 
phase of the YOLO -V5 framework, a total of 59,150 class -wise 
data samples wer e utilized, along with their corresponding class 
labels  
 
 
 
 
 
 
 
 
 
 
 
 
 
Fig. 13.  Depicts the respective class -wise training samples 
distributions . 
B. Data Annotation and Augmentation  
The overall data annotations were performed manually using 
an open -source bounding box -based annotations tool LabelImg 
[31] for all the thermal classes in our study. Annotations are 
stored  in YOLO format as text files. During the training phase 
all the YoloV5 network variations which include small, 
medium, large, and x -large networks have been trained to detect 
and classify six different classes in different environmental 
conditions.  
Large -scale datasets are considered a vital requirement for 
achieving optimal training results using deep learning 
architectures. Without the need of gathering new data, data 
augmentation allows us to significantly improve the diversity 
of data available that c an be effectively used for training the 
DNN models. In the proposed study we have incorporated a 
variety of data augmentation techniques which involve 
cropping, flipping, rotation, shearing, translation, mosaic 
transformation for an optimum training of all  the network 
variants of the YOLO -V5 framework . 
A. C. Training Results  
As discussed in subsection A of section IV all the networks 
are trained using the combination of public as well as the locally 
gathered dataset. Training data from public datasets are 
included from four different datasets which include FLIR [ 7], 
OST [ 2], CVC [ 19], and KAIST [ 5] datasets. Secondly, we have 
used thermal frames acquired from the locally gathered video 
sets using both M1 and M2 methods. The training process is 
performed on a server -grade machine with XEON E5 -1650 v4 
3.60 GHz processor, 64 GB of ram, and  equipped with 
GEFORCE RTX 2080  Ti graphical processing unit. It comes 
with 12 GB of dedicated graphical memory, memory bandwidth 
of 616 GB/second, and 4352 cuda cores . During  the training 
phase the batch size is fixed to 32 and as an optimizer, both 
stochastic gradient descent  (SGD) and ADAM optimizer were 
used. However, we were unable to achieve satisfactory training 
results using ADAM optimizer as compared to SGD thus 
select ed SGD optimizer for training purposes. Table  III shows 
the performance evaluation of all the trained models in the form 
Training Data                           
Locally Acquired + Public 
Datasets  
Classes  
 
Car, Pole, Bike, Bicycle, Person, Bus
  
Data Annotation  
Data Augmentation Techniques  
 
Flipping, Cropping, Shearing, Rotation, Translation, Mosaic  
YOLO v5 Network Variants  
 
Small  
(7.3 million 
parameters ) 
Medium  
(21 million 
parameters ) 
Large  
(47 million 
parameters ) 
X-Large 
(87.7m illio
n 
parameters ) 
Inference Testing  (On both public and local dataset)  
 
GPU, Nvidia -Jetson, Nvidia -Xavier  
7 
 
 of mean average precision (mAP), recall rate, precision, and 
losses.  
TABLE III 
TRAINING RESULTS  
Optimizer: SGD  (best model  *)  
Network  P % R% mAP
% Box 
Loss  Object 
Loss  Classific
ation 
Loss  
Small  75.5
8 65.75  70.71 0.03
2 0.034  0.0017  
Medium  71.0
6 64.74  65.34 0.02
7   0.030  0.0013  
Large  * 82.2
9 68.67  71.8 0.02
5   0.0287  0.0011  
X-Large  74.2
3 65.03  64.94  0.02
5  0.0270  0.0010  
 
By analy sing Table III, it can be observed that the large model 
performed significantly better when compared to other models 
with an overall precision of 82.29%, recall rate of 68.67%, and 
mean average precision of 71.8% mAP . Fig.  14 shows the 
graph result s of yolo-v5 large model.  The figure visualizes 
obtained PR -curve, box loss, object loss, and classification loss. 
During the tra ining process, the X -large model consumes the 
maximum amount of hardware resources with the largest 
training time as compared to other network variants with overall 
GPU usage of 9.78 GB and a total training time of 14 hours . 
Fig. 15 shows the overall GPU m emory usage, GPU power 
required in percentages, and GPU temperature in centigrade 
scale while training the largest x -large network variant of yolo -
v5 model.  
 
 
 
 
 
 
Fig. 14.  Training results of YOLO -v5 large model  using SGD  
optimizer . 
V. VALIDATION RESULTS ON GPU  AND EDGE DEVICES  
This section will demonstrate the object detection validation 
results on GPU as well as on two different embedded boards.  
A. Testing Methodology and Overall Test Data  
    In this research study , we have used three different testing 
approaches which include the conventional test -time method 
with no augmentation (NA), test -time augmentation (TTA), and 
test-time with model ensembling (ME). TTA is an extensive 
application of data augmentation applied  to the test dataset. It 
performs by creating multiple augmented copies of each image 
in the test set, having the model make a prediction for each, then 
returning an ensemble of those predictions.  However, since the 
test dataset is enlarged  with a new set of augmented images the   
 
 
 
Fig. 15. GPU resource utilization during the training process of 
x-large network, a) 85% (9.78 GB) of GPU memory utilized, 
(b) 90% (585 watts) of GPU power required and, (c) 68 C of 
GPU temperature with the maximum rating of 89 C.   
 
overall inference time also increases as compared to NA which 
is one of the downsides of this approach . TTME or ensemble 
learning refers to as using multiple trained networks at the same 
time in a parallel  manner to produce one optimal predictive  
inference model  [35]. In this  study, we have tested the 
performance of individually trained variants of the Yolo -V5 
framework and selected the best combination of models which 
in turn helps in achieving better  validation results.  
After training all the networks variants of yolo -v5, the 
performance of each model is cross -validated on a 
comprehensive set of test data  selected from the public as well 
as locally gathered thermal data. Table IV provides the numeric 
data distribution of the overall validation set.  
TABLE IV 
TEST DATASET  
Test Dataset Attributes  
Frames Used  
Public 
dataset  OST  CVC -09 KAIS
T FLI
R Total No 
frames  
 
 
50 5360 
(day +  
night -
time)  149 130 5,689  
Local 
dataset  Method (M1)  Method (M2)  Total No 
frames  
a 
b 
c 100 
 80 
 60 
 40 
 20 
 60  120  20 minutes  GPU Memory Usage  
GPU Power Usage  
GPU Temperature   20  60  120 minutes   20  40  60  80  100 
minutes   20  60  120  10  20  30  40  50  60 8 
 
 
 
8,820 16,560  25,380  
 Total: 31,069  
B. Inference Results Using YOLO Network Variants  
In the first phase, we have run the  rigorous inference test on 
GPU as well as Edge -GPU platforms on our test data using the 
newly trained networks variants of yolo framework. The overall 
test data is consisting  of nearly ≈ 31,000 thermal frames. Fig. 1 6 
shows the inference results on 9 diffe rent thermal frames 
selected from both public as well as locally acquired data. These 
frames have data complications such as multiple class objects, 
occlusion, overlapping classes, scale variation, and varying 
environmental conditions . The complete inference results are 
available on our local repos itory ( https://bit.ly/3lfvxhd ).  
In the second phase , we have run t he combination of different 
models in a parallel manner using the model ensembling 
approach to output one optimal predictive engine which can be 
further used to run the inference test on the validation set. The 
different combination of these models is show n in Table V  
respectively where 1 indicates that model is in active state and 
0 means model is in a non -active state.  
TABLE V 
MODEL ENSEMBLING  
Model Combinations  
N
o  Small   Medium  Large  X-Large  Combination  
State 1 (active) or 0 (not active)  
1 1 1 0 0 A0 
2 1 0 1 0 A1 
3 1 0 0 1 A2 
4 0 1 1 0 A3 
5 0 0 1 1 A4 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Fig. 16 . Inference results on nine different frames selected 
from test data.  
 
With the model ensembling method small and large models 
(A1) turn out to best model combination in terms of achieving 
the best mAP, recall, and relatively less amount of inference 
time per frame thus producing optimal validation results. These 
results are examined in further parts of this section.  Fig. 1 7 
shows the inference results using A1 model ensembling engine 
on three  different thermal frames s elected  from the test data.   
 
 
 
 
 
 
Fig. 17.  Inference results on three different frames using 
model ensembling.  
C. Quantitative Validation Results on GPU  
The third part of the testing phase shows the quantitative 
numerical results of all the trained models on GPU. To better 
analy se and validate the overall performance  for all the trained 
models  on test data, relatively a smaller set of test images has 
been selected from the overall test  set. For this purpose, a subset  
of 402 thermal frames  is selected to compute all the evaluation 
metrics. The selected images consist of different roadside 
objects such as pedestrians, cars and buses under different 
illum ination and environmental conditions, time of day, and 
distance from the camera. The objects are either far -field 
(between 11 -18 meters) , mid -field (between 7 -10 meters) or 
near-field (between 3 -6 meters) from the camera. Fig. 18 shows 
selected views from the test data for quick reference of the 
reader.  
 
 
 
 
 
 
 
 
Fig. 18.  Test data samples with the object at varying distances 
from the camera, (a) near -field distance, (b) mid -field distance, 
(c) far -field distance.  
 
The performance evaluation of each model is computed 
using four different metrics which include recall, preci sion, 
mean average precision (mAP), and frames per second rate 
(FPS). Table VI shows all the quantitative validation results on 
GPU. During the testing phase batch size is fixed to 8. Also, 
three different testing configuration is selected thus having 
separate confidence threshold values and the intersection of 
union values at each validation phase. Confidence threshold 
defines the minimum threshold value, or in other words, it is the 
minimum confidence score above which we consider a 
prediction as true. If  it’s below the threshold value, we consider 
the prediction as “no”. The last row of Table VI  shows the best 
ME results using A1 configuration from Table V  with a selected 
confidence threshold of 0.2 and IoU threshold of 0.4.  
TABLE VI 
QUANTITATIVE RESULTS ON GPU  
Platform: GPU  
Inference image size: 800 x 800  
Confidence Threshold: 0.4, IoU Threshold: 0.6  
No Augmentation (NA)  Test-time Augmentation 
(TTA)  
Network  P 
% R
% mA
P% FPS P 
% R
% mA
P% FPS 
a b c camera  camera  camera  person  person  car 9 
 
 Small  72 46 43 79 76 48 50 45 
Medium  73 54 49 53 76 58 57 26 
Large   75 56 52 34 77 63 60 16 
X-Large  74 53 49 20 71 59 55 10 
Confidence Threshold: 0.2, IoU Threshold: 0.4  
NA TTA 
Small  66 50 47 82 64 55 52 45 
Medium  66 57 51 53 77 58 59 27 
Large   71 61 56 35 78 63 63 16 
X-Large  70 54 50 21 68 62 56 10 
Confidence Threshold: 0.1, IoU Threshold: 0.2  
NA TTA 
Small  65 52 48 81 65 53 53 45 
Medium  69 54 51 53 77 58 59 26 
Large   73 61 57 34 79 63 63 16 
X-Large  71 54 52 21 69 62 57 10 
Confidence Threshold: 0.2, IoU Threshold: 0.4  
Model Ensembling (ME)  
A = Small  
B = Large  
Comb: A1  
 --- --- --- --- 77 66 65 25 
 
D. Quantitative Validation Results on Edge -GPU Devices  
This section will review the quantitative validation results on 
two different Edge -GPU platforms (Jetson Nano & Jetson 
Xavier NX).  It is pertinent to mention that Jetson Xavier NX 
development kit embeds more computational power in terms of 
GPU, CPU, and memory as compared to Nvidia Jetson Nano. 
Table  VII shows the specification comparison of both boards.  
TABLE VII 
HARDWARE SPECIFICAT ION COMPARISON  
Hardware specification comparison o f Nvidia Jetson Nano and 
Nvidia Jetson Xavier NX 
Board Jetson Nano [23] Jetson Xavier N X [25] 
CPU  Quad -Core  ARM® 
Cortex® -A57 MPCore, 
2 MB L2,  Maximum 
Operating 
Frequency:  1.43 GHz  6-core NVIDIA Carmel 
ARM®v8.2 64 -bit 
CPU, 6 MB L2 + 4 MB 
L3, Maximum 
Operating 
Frequency:  1.9 GHz  
GPU  128-
core Maxwell  GPU, 
512 GFLOPS (FP16), 
Maximum Operating 
Frequency: 921 MHz  384 CUDA® cores + 
48 Tensor 
cores  Volta  GPU, 21 
TOPS, Maximum 
Operating Frequency: 
1100 MHz  
RAM  4 GB  64-bit LPDDR4 
@ 1600MHz | 25.6 
GB/s  8 GB  128-bit 
LPDDR4x @ 
1600MHz | 51.2GB/s  
On module 
Storage  16 GB  eMMC 5.1 Flash Storage, Bus Width: 8 -bit, 
Maximum Bus Frequency: 200 MHz (HS400)  
Thermal 
Design 
Power  5W – 10W  10W – 15W  
AI 
Performance  0.5 TFLOPS (FP16)  6 TFLOPS (FP16)  
21 TOPS (INT8)  
 
On Jetson Nano we have validated the performance of the 
small version only whereas on Jetson Xavier NX we have 
evaluated the performance of smaller and medium versions of 
models due to the memory limitations and constrained 
hardware resources on these boar ds. During the testing phase, 
we have selected the highest power modes on both boards to 
provide the utmost efficiency thus utilizing maximum hardware 
resources. For instance, on Nvidia Xavier board NX we have selected ‘Mode Id: 2’ which means the board is  operating in 15 -
watt power mode with all the six cores active with a maximal 
CPU frequency of 1.4 gigahertz and GPU frequency of 1.1  
gigahertz. Similarly, on Nvidia Jetson Nano all the four CPU 
cores were utilized with overall power utilization of 5 watts . 
Table  VIII shows the quantitative validation results  on ARM 
processor based embedded boards  
TABLE VIII 
QUANTITATIVE RESULTS ON EDGE PLATFORMS  
Platform: N vidia Jetson Nano  
Inference image size: 128 x 128  
Confidence Threshold: 0.4, IoU Threshold: 0.6  
NA TTA  
 P 
% R
% mA
P% FPS P 
% R
% mA
P% FPS 
 
Small  75 44 45 3 77 47 49 1 
Confidence Threshold: 0.2, IoU Threshold: 0.4  
NA TTA  
Small  75 44 47 3 71 51 51 1 
Confidence Threshold: 0.1, IoU Threshold: 0.2  
NA TTA  
Small 66 47 48 2 73 50 52 1 
Platform: N vidia Jetson  Xavier NX  
Inference image size: 128 x 128  
Confidence Threshold: 0.4, IoU Threshold: 0.6  
NA TTA  
Small  75 44 45 18 77 47 49 10 
Med 76 53 50 12 79 50 52 6 
Confidence Threshold: 0.2, IoU Threshold: 0.4  
NA TTA 
Small  75 44 47 19 71 51 51 10 
Med 76 52 53 12 73 54 53 6 
Confidence Threshold: 0.1, IoU Threshold: 0.2  
NA TTA 
Small  66 47 48 18 73 50 52 10 
Med 76 51 52 12 81 49 53 6 
E. Real-time Hardware Feasibility Testing  
While running these tests we closely monitor the temperature 
ratings of different hardware peripherals on both Edge -GPU 
platforms. It is done to prevent the overheating effect which can 
damage the onboard processor or effect the overall operational 
capabil ity of the system. In the case of Nvidia Jetson Nano, a 
cooling fan was mounted on top of the processor heatsink to 
reduce the overheating effect as shown in Fig. 1 9. 
 
 
Fig. 19.  External 5 -volt fan unit mounted on Nvidia Jetson 
Nano processor heatsink to avoid onboard overheating effect 
while running the inference testing.  
The temperature ratings of various hardware peripherals are 
monitored using eight different on -die thermal s ensors and one 
on-die thermal diode. These temperature monitors are referred 
to as CPU -Thermal, GPU -Thermal, Memory -Thermal, and 
PLL-Thermal (part thermal zone). External fans help us in 
External fan  10 
 
 
reducing the temperature rating of various hardware peripherals 
drast ically as compared to without mounting the fan. Fig. 20 
shows the temperature rating difference of onboard thermal 
sensors while running the smaller version of the model on 
Nvidia Jetson Nano without and with mounting the external 
cooling fan.  
 
 
 
 
 
Fig. 20.  Temperature rating difference of different onboard 
hardware peripherals on Jetson Nano (a) without fan: A0 
thermal zone = 65.50 C, CPU = 55 C, GPU = 52 C, PLL: 53.50, 
overall thermal temperature = 53.50 C, (b) with external fan: 
A0 therm al zone = 45.50 C, CPU = 33 C, GPU = 33 C, PLL: 
33, overall thermal temperature = 32.75 C.  
 
It can be examined from Fig. 20 that by mounting an external 
cooling fan the temperature rating of  various onboard 
peripheral on Jetson Nano was reduced by nearly 30% thus 
allowing us to operate the board at its maximum capacity for 
rigorous model testing. Fig. 21 shows the Nvidia Jetson running 
at its full pace  (with an external fan)  such that all the four cores 
running at their maximum limit (100% capacity) while ru nning 
the quantitative  and inference test by deploying the smaller 
network variant of the yolo -v5 framework.   
 
Fig. 21.  Nvidia Jetson Nano running at MAXN power mode 
with all the cores running at their maximum capacity while 
running t he inference test and quantitative validation test.  
Fig. 22 shows the temperature rating difference of onboard 
thermal sensors while running the smaller version of the model 
on Nvidia Jetson Xavier NX board. Whereas Fig. 23 shows the 
CPU and GPU usage while running the smaller variant of YOLO -V5 framework for quantitative validation and inference 
test on Nvidia Xavier NX development kit.  
 
Fig. 22.  Temperature rating of different onboard hardware 
peripherals on Jetson Xavier NX (a) A0 thermal zone = 41.50 
C, AUX: 42.5 C, CPU = 44 C, GPU = 42 C, overall thermal 
temperature = 42.80 C,  
 
 
 
 
 
Fig. 23.  Nvidia Jetson Xavier running at 15 -watt 6 core power 
mode, (a) all the CPU cores running at its maximum capacity 
while running the quantitative validation test, (b) 69% GPU 
utilization while running the inference test with an image size 
of 128 x 128.  
VI. MODEL PERFORMANCE OPTIMIZATION (S)  
This section will mainly aim at further model optimization 
using TensorRT [ 33] inference accelerator tool. The prime 
reason for this is to further increase the FPS rate  for real -time 
evaluation and on -board feasibility te sting on edge devices.  
Secondly, it helps in saving onboard memory footprints on the 
target device by performing various optimization methods.  
TensorRT  [33] works by performing five modes of 
optimization  methods  for increasing the throughput of deep  
neural networks . In the first step, it maximizes throughput by 
quantizing models to 8 -bit integer data type or FP16 precision 
while preserving the model accuracy. This method significantly 
(a) 
(b) 
(a) 
(b) 11 
 
 
reduces the model size since it is transformed from originally 
FP32 to FP16 version. In the next step, it uses layer and tensor 
fusion  techniques to further optimize the usage of onboard GPU 
memory. The third step includes perform ing kernel auto -tuning.  
It is the most important step where the TensorRT engine 
shortlists  the best network layers,  and optimal batch size based 
on the target GPU  hardware. In the second last step, it 
minimizes memory footprint s and re -uses memory by 
distributing memory to tensor only for the period of its usage.  
In the last steps, it processes m ultiple input streams in parallel 
and finally optimizes neural networks periodically with 
dynamically generated kernels [ 33]. 
In the proposed research work we have deployed a smaller 
variant of yolo -v5 using TensorRT inference accelerator on 
both edge plat forms Nvidia Jetson Nano and Nvidia Jetson 
Xavier NX development boards to further excel the 
performance of the trained model.  It produces faster inference 
time thus increasing the FPS on thermal data which in turn helps 
us in building an effective real -time forward sensing system for 
ADAS embedded  applications.  Fig. 24 depicts the block 
diagram representation of deployment phase TensorRT 
inference accelerator on embedded platforms.  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Fig. 24. Overall block diagram representation of deployment 
and running TensorRT inference accelerator on two different 
embedded platforms.  
 
Table  IX shows the overall inference time along with FPS 
rate on thermal test data using TensorRT  run-time engine.  By 
analyzing the results from Table. IX we can deduce that 
TensorRT API supports in boosting the overall FPS rate on 
ARM -based embedded platforms by nearly 3.5 times as 
compared to the FPS rate achieved by running the non -
optimized smaller variant on Nvidia Jetson Nano and Nvidia 
Jetson Xavier boards. The same is demonstrated via graphical 
chart results in Fig. 2 5. 
TABLE IX 
TENSOR RT INFERENCE ACCELERATOR RESULTS  
FPS on Nvidia Jetson Nano and Nvidia Jetson Xavier NX 
Board Nvidia Jetson Nano  Nvidia Jetson Xavier 
NX Test Data  402 images with the resolution of 128x128  
Overall 
inference time  35,090 milliseconds 
≈ 35.1 seconds  6,675 milliseconds ≈ 
 6.7 seconds  
PS 35.1 sec / 402 frame s 
= 0.087 sec/frame  
 
FPS: 1 sec / 0.087 = 
11.49 ≈ 11 fps  6.7 sec / 402 frames = 
0.0166 sec/frame  
 
FPS:  1 sec / 0.0166 = 
60.24 ≈ 60 fps  
 
 
 
Fig. 25.  FPS increment rate of nearly 3.5 times on Jetson Nano 
and Jetson Xavier NX embedded boards using the TensorRT 
built optimized inference engine.  
 
Fig. 2 6 shows the  thermal object detection inference results on 
six different thermal frames from the public as well as locally 
acquired test data produced  through the neural accelerator.  
   
 
 
Fig. 2 6. Inference results using TensorRT neural accelerator, 
(a) Object detection results on public data, (b) Object Detection 
results on locally acquired thermal frames.  
DISCUSSION / ANALYSIS  
This section  will review the training and testing performance 
of all YO LO-V5 framework model variants.  
• During the training phase, the large YOLO v5 network 
outperforms other network variants scoring the highest 
precision of 82.29 % and a mean average precision (mAP) 
score of 71.8 %.  
• Although  the large network variant performed  significantly 
better  during the training phase , the small network variant 020406080
N V ID I A  J E T S O N  N A N O N V ID I A  J E T S O N  
X A V IE R  N XF P S  C O M P A R A S IO N  U S IN G  O P T IM IZ E D  
A N D  N O N -O P T IM I Z E D  V E R S IO N  O F  
S M A LL N E T W O R K  V A R IA N T
Non-optimized version TensorRT optimized version
Smal ler variant 
training on 
thermal data 
GPU  
Trained 
weights  
TensorRT Inference Engine  
1. Nvidia Jetson Nano  
 
2. Nvidia Jetson Xavier NX  
 
Jetson optimized  
runtime engine  
Xavier optimized  
runtime engine  
Test data  
(a) 
b 12 
 
 also performed well with an overall precision of 75.58 % and 
mAP of 70.71%.  Also, it gains a higher FPS rate on GPU 
during the testing phase as compared to the large model.  
Fig. 2 7 summarizes the quantitative performance 
comparison of small and large network variants of yolo 
framework.  
 
Fig. 2 7. Quantitative metrics comparison of small and large 
network variants  
• Due to the smaller number of model parameters as 
compared to larger network variant ( 7.3M Vs 47M model 
parameters)  and faster FPS rate on GPU during the testing 
phase  as shown in  Fig. 26  this model is shortlisted for 
validation and deployment purposes on both the edge 
embedded platforms Nvidia Jetson Nano and Nvidia  Jetson 
Xavier NX kits.  
• During the testing phase, it was  noticed that by reducing the 
confidence threshold from 0.4 to 0.1 and the IoU threshold 
from 0.6 to 0.2 in three  stepwise intervals,  the model's mAP 
and recall rates increased significantly, but the precision 
level decreas es. However, the FPS rate remains effectively 
constant in most of the trained model cases.  
• TTA methods achieved improved testing results when 
compared to the NA method however the main drawback of 
this method is that the FPS rate dro ps substantially which is 
not suitable for real -time deployments. To overcome this 
problem a model ensembling (ME) based inference engine 
is proposed. Table IV shows the ME results by running 
large -small model in parallel configuration with a 
confidence threshold  of 0.2, and an IoU Threshold  of 0.4. 
The ensembling engine attains an overall mAP of 66% with 
25 frames per second.  
• When comparing the individual hardware resources of both 
the edge platforms (NVidia Jetson Nano and Jetson Xavier), 
Xavier is computationally more powerful than the Jetson 
Nano. Note that d ue to memory limitations and the lower 
computational power of the Jetson  only the small network 
variant was evaluated on  the Jetson Nano, whereas both the 
smaller and medium  network variants we re evaluated on  the 
Jetson Xavier NX.  
• It was observed that t hroughout the testing  phase , it was 
important to keep a close eye on the operational temperature 
ratings of different onboard thermal sensors  to avoid 
overheating, which might damage the onboard components or affect the system's typical operational performance.  
Active cooling fans were used on both boards during testing, 
and both ran at close to their rated temperature limits.   
• This study also included model optimization using 
TensorRT [33] inference accelerator tool. It was  determine d 
that TensorRT leads  to an approximate increase of FPS rate 
by a factor of 3.5 when compared to the non-optimized 
smaller variant of yolo -v5 on Nvi dia Jetson Nano and 
Nvidia Jetson Xavier devices.  
• After performing model optimization, the Nvidia Jetson 
produced 11 FPS and Nvidia Jetson Xavier achieved 60 FPS  
on test data . 
CONCLUSION  
Thermal imaging provides superior and effective results in 
challeng ing environments such that in low lighting scenarios 
and has aggregate immunity to visual limitations  thus making it 
an optimal solution for intelligent and safer vehicular systems . 
In this study, we presented a new benchmark thermal dataset 
that comprises  over 35K distinct frames recorded, analyzed, 
and open -sourced in challenging weather and environmental 
conditions utilizing a low -cost yet reliable uncooled LWIR 
thermal camera. All the YOLO v5 network variants were  
trained using locally gathered data as well as four different 
publicly available datasets. The performance of trained 
networks is analysed  on both GPU as well as ARM processor -
based edge devices for onboard automotive sensor suite 
feasibility testing . On edge devices , the small and medium 
network edition of YOLO is deployed and tested  due to certain 
memory limitations and less computational power of these 
boards . Lastly, we further optimized the smaller network 
variant using TensorRT inference accelerator to explicitly 
increase the FPS on edge devices. This allowed the system to 
achieve  11 frames per second  on jetson nano , while the Nvidia 
Jetson Xavier delivered  a significantly higher performance of  
60 frames per second . These results  validate the potentia l for 
thermal imaging as a core component  of ADAS systems for 
intelligent vehicles .  
As the future directions, the system's performance can be 
further enhanced by porting the trained networks on more 
advanced and powerful edge devices thus tailoring it for real -
time onboard deployments. Moreover, the current system 
focuses on object recognition, but it can be enhanced to 
incorporate image segmentation, road and lane detection, traffic 
signal and road signs classification, and object tracking  for 
providing comprehensive driver assistanc e.  
ACKNOWLEDGMENT  
The authors would like to acknowledge Cosmin Rotariu from 
Xperi -Ireland and the rest of the team members for providing 
the support in preparing the data accusation setup and helping 
throughout  in data collection and Quentin Noir from Lynred 
France for giving their feedback. Moreover, the authors would 
like to acknowledge the contributors of all the public datasets 
for providing the image resources to carry out this research 
work and ultralytics  for sharing the YOLO -V5 Pytorch version.  020406080100
Small LargePerformance Comparsion of Small vs Large Model
Model Parameters mAP Recall Precision FPS13 
 
 REFERENCES  
[1] Lynred France,  https://www.lynred.com  /  (Last accessed on 20th July 
2021) . 
[2] Davis, James W., and Mark A. Keck. "A two -stage template approach to 
person detection in thermal imagery." in proceedings of 2005  Seventh 
IEEE Workshops on Applications of Computer Vision 
(WACV/MOTION'05) -Volume 1 , vol. 1,  IEEE, 2005 , pp. 364 -369. 
https://doi.org/10.1109/ACVMOT.2005.14 . 
[3] A. Torabi, G. Massé, and G. A. Bilodeau, “An iterative integrated 
framework for thermal -visible image registration, sensor fusion, and 
people tracking for video su rveillance applications,” Comput. Vis. Image 
Underst., vol. 116, no. 2, 2012, doi: 10.1016/j.cviu.2011.10.006.  
[4] Z. Wu, N. Fuller, D. Theriault, and M. Betke, “A thermal infrared video 
benchmark for visual analysis,” 2014, doi: 10.1109/CVPRW.2014.39.  
http://www.vcipl.okstate.edu/otcbvs/bench/  
[5] Hwang, S., Park, J., Kim, N., Choi, Y., & Kweon, I. S. (n.d.). 
Multispectral Pedestrian Detection: Benchmark Dataset and Baseline . 
http://rcv.kaist.ac.kr/multisp ectral -pedestrian/  
[6] Z. Xu, J. Zhuang, Q. Liu, J. Zhou, and S. Peng, “Benchmarking a large -
scale FIR dataset for on -road pedestrian detection,” Infrared Phys. 
Technol. , vol. 96, 2019, doi: 10.1016/j.infrared.2018.11.007.  
[7] FLIR Thermal Dataset . [Online] Available: 
‘https://www.flir.com/oem/adas/adas -dataset -form/’ , (Last accessed on 
22nd July 2021).  
[8] R. R. Ziyatdinov and R. A. Biktimirov, “Automated system of 
recognition of road signs for ADAS systems,” in IOP Conference Series: 
Materials Scienc e and Engineering , 2018, vol. 412, no. 1, doi: 
10.1088/1757 -899X/412/1/012081.  
[9] O. Töro, T. Bécsi, and S. Aradi, “Performance Evaluation of a Bernoulli 
filter Based Multi -Vehicle Cooperative Object Detection,” in 
Transportation Research Procedia , 2017, vol.  27, doi: 
10.1016/j.trpro.2017.12.042.  
[10] G. Yan, M. Yu, Y. Yu, and L. Fan, “Real -time vehicle detection using 
histograms of oriented gradients and AdaBoost classification,” Optik 
(Stuttg). , vol. 127, no. 19, 2016, doi: 10.1016/j.ijleo.2016.05.092.  
[11] A. M. Ali , W. I. Eltarhouni, and K. A. Bozed, “On -road vehicle detection 
using support vector machine and decision tree classifications,” 2020, 
doi: 10.1145/3410352.3410803.  
[12] V. Ghenescu, E. Barnoviciu, S. V. Carata, M. Ghenescu, R. Mihaescu, 
and M. Chindea, “Object  recognition on long range thermal image using 
state of the art dnn,” 2018, doi: 10.1109/ROLCG.2018.8572026.  
[13] A. Narayanan, R. Darshan Kumar, R. Roselinkiruba, and T. Sree 
Sharmila, “Study and Analysis of Pedestrian Detection in Thermal 
Images Using YOLO an d SVM,” 2021, doi: 
10.1109/WiSPNET51692.2021.9419443.  
[14] R. Kalita, A. K. Talukdar, and K. Kumar Sarma, “Real -Time Human 
Detection with Thermal Camera Feed using YOLOv3,” 2020, doi: 
10.1109/INDICON49873.2020.9342089.  
[15] A. Corovic, V. Ilic, S. Duric, M. Marijan, and B. Pavkovic, “The Real -
Time Detection of Traffic Participants Using YOLO Algorithm,” 2018, 
doi: 10.1109/TELFOR.2018.8611986.  
[16] M. Ivašić -Kos, M. Krišto, and M. Pobar, “Human detection in thermal 
imaging using YOLO,” in ACM International Conferen ce Proceeding 
Series , 2019, vol. Part F148262, doi: 10.1145/3323933.3324076.  
[17] X. Han, J. Chang, and K. Wang, “Real -time object detection based on 
YOLO -v2 for tiny vehicle object,” in Procedia Computer Science , 2021, 
vol. 183, doi: 10.1016/j.procs.2021.02.03 1. 
[18] V. Paidi, H. Fleyeh, and R. G. Nyberg, “Deep learning -based vehicle 
occupancy detection in an open parking lot using thermal camera,” IET 
Intell. Transp. Syst. , vol. 14, no. 10, 2020, doi: 10.1049/iet -
its.2019.0468.  
[19] CVC -09 FIR Sequence Pedestrian Dataset. Available 
online:  http://adas.cvc.uab.es/elektra/enigma -portfolio/item -1/ (Last 
accessed on  22nd June 2021). 
[20] S. Houben, J. Stallkamp, J. Salmen, M. Schlipsing, and C. Igel, 
“Detection of traffic signs in real -world images: The German traffic sign 
detection benchmark,” 2013, doi: 10.1109/IJCNN.2013.6706807.  
[21] L. Zhang, J. Tan, D. Han, and H. Zhu, “From machine learning to deep 
learning: progress in machine intelligence for rational drug discovery,” 
Drug Discovery Today , vol. 22, no. 11. 2017, doi: 
10.1016/j.drudis.2017.08.010.  
[22] Pretrained object detection and classification models , [Online] 
Available, https://pytorc h.org/serve/model_zoo.html , (Last accessed on 
29th August 2021).  [23] Nvidia Jetson Nano,  [Online] Available : 
https://developer.nvidia.com/embedded/jetson -nano -developer -kit, 
(Last accessed on 14th J une 2021).  
[24] M. Kristo, M. Ivasic -Kos, and M. Pobar, “Thermal Ob ject Detection in 
Difficult Weather Conditions Using YOLO,” IEEE Access , vol. 8, 2020, 
doi: 10.1109/ACCESS.2020.3007481.  
[25] Nvidia Jetson Xavier NX Development kit , [Online] Available : 
https://developer.nvidia.com/embedded/jetson -xavier -nx-devkit , (Last 
accessed on 14th J une 2021).  
[26] A. Farouk Khalifa, E. Badr, and H. N. Elmahdy, “A survey on human 
detection surveillance systems for Raspberry Pi,” Image Vis. Comput. , 
vol. 85, 2019, doi: 10 .1016/j.imavis.2019.02.010.  
[27] D. S. Breland, S. B. Skriubakken, A. Dayal, A. Jha, P. K. Yalavarthy, 
and L. R. Cenkeramaddi, “Deep Learning -Based Sign Language Digits 
Recognition from Thermal Images with Edge Computing System,” IEEE 
Sens. J. , vol. 21, no. 9, 2021, doi: 10.1109/JSEN.2021.3061608.  
[28] Y. Liu, S. Cao, P. Lasang, and S. Shen, “Modular Lightweight Network 
for Road Object Detection Using a Feature Fusion Approach,” IEEE 
Trans. Syst. Man, Cybern. Syst. , vol. 51, no. 8, 2021, doi: 
10.1109/TSMC.2019.294505 3. 
[29] J. Fritsch, T. Kuhnl, and A. Geiger, “A new performance measure and 
evaluation benchmark for road detection algorithms,” 2013, doi: 
10.1109/ITSC.2013.6728473.  
[Online] Available, http://www.cvlibs.n et/datasets/kitti/  (Last accessed 
on 14th June 2021)  
[30] G. Balasekaran, S. Jayakumar, and R. Pérez de Prado, “An intelligent 
task scheduling mechanism for autonomous vehicles via deep learning,” 
Energies , vol. 14, no. 6, 2021, doi: 10.3390/en14061788.  
[31] LabelIm g object annotation tool,  [Online] Available  
https://github.com/tzutalin/labelImg , (Last accessed on 20th July 2021) . 
[32] Heliaus European Union Project, https://www.heliaus.eu/   (Last accessed 
on 20th February 2021) . 
[33] Nvidia TensorRT for developers, [Online] Available,  
https://developer.nvidia.com/tensorrt, (Last accessed on 14th February 
2021).  
[34] J. de Vries, “Image proc essing and noise reduction techniques for 
thermographic images from large -scale industrial fires,” 2014.  
[35] Ganaie, M. A., and Minghui Hu. "Ensemble deep learning: A 
review."  arXiv preprint repository  arXiv:2104.02395  (2021) . 
[36] Saragaglia, Amaury, and Alain Durand. "Method of infrared image 
processing for non -uniformity correction." U.S. Patent No. 10,015,425. 
3 Jul. 2018.  
 
Muhammad Ali Farooq  received his BE 
degree in electronic engineering from  
IQRA University  in 2012 and his MS 
degree in electrical  control engineering 
from the National University  of Sciences 
and Technology (NUST) in 2017.   He is 
currently pursuing the  Ph.D. degree at the 
National University of  Ireland  Galway  
(NUIG) . His research interests include 
machine vision , computer vision, video analytics, and sensor 
fusion. He has won the prestigious H2020 European Union  
(EU)  scholarship and  currently working at NUIG as one of the 
consortium partners in the H eliaus  (thermal v ision augmented 
awarenes s) project  funded by EU.  
Waseem Shariff  received  his B.E degree 
in computer science from Nagarjuna 
College of Engineering and Technology 
(NCET) in 2019 and his M.S. degree in 
computer  science,  specializing  in artificial  
intelligence  from National  University  of 
Ireland  Galway  (NUIG)  in 2020.  He is 
working  as research  assistant  at National  
University  of Ireland  Galway  (NUIG).  He 
is associated  with Heliaus  (thermal v ision augmented 
awarenes s) project.  He is also allied  with FotoNation/Xperi  
14 
 
 research  team.  His research  interests  include  machine  learning  
utilizing  deep  neural  networks  for computer  vision  applications,  
including  working  with synthetic  data,  thermal  data,  and RGB . 
 
Peter Corcoran  (Fellow, IEEE) holds a 
Personal  Chair in Electronic Engineering at 
the College of Science and Engineering, 
National University of Ireland Galway  
(NUIG) . He was the Co -Founder in several 
start-up companies, notably FotoNation, 
now the Imaging Division of  Xperi 
Corporation. He has more th an 600 cited 
technical publications and patents, more than 120 peer -
reviewed journal articles, 160 international conference papers, 
and a co -inventor on more than 300 granted U.S. patents. He is 
an IEEE Fellow recognized for his contributions to digital 
camera technologies, notably in -camera red -eye correction and 
facial detection. He is a member of the IEEE Consumer 
Technology Society for more than 25 years and the Founding 
Editor of IEEE Consumer Electronics Magazine.