Edge Deployable Lightweight Framework for Real-Time Driver Safety System with Integrated Lane Departure and Traffic Sign Recognition
Authors: Hassaan Ali Qureshi, Danish Hussain, Ali Hassan, Aon Safdar, Anas Bin Aqeel, Muhammad Rafiq, Zohaib Riaz, Farhan Hussain
Organizations: Department of Mechatronics Engineering, College of Electrical and Mechanical Engineering, National University of Sciences and Technology, Islamabad 44000, Pakistan · Department of Robotics and Artificial Intelligence, School of Mechanical and Manufacturing Engineering, National University of Sciences and Technology, Islamabad 44000, Pakistan · Department of Computer and Software Engineering, College of Electrical and Mechanical Engineering, National University of Sciences and Technology, Islamabad 44000, Pakistan · School of Computer Science, University College Dublin, Dublin, Ireland · School of Energy Science and Engineering, Chongqing Research Institute of Harbin Institute of Technology, Chongqing 150001, PR China
Lane departure and traffic-sign negligence are leading contributors to road accidents, particularly in regions with non-standardized signage and limited annotated data. Existing Advanced Driver Assistance Systems (ADAS) solutions are often developed as disjointed submodules, are rarely validated for real-time operation, and struggle to balance accuracy with low-latency inference for embedded deployment. We address these gaps with a unified, edge-deployable ADAS framework that simultaneously performs lane-departure warning (LDW) and traffic-sign detection and recognition (TSD/TSR), and validate the proposed system on automotive-class devices. Our framework combines a lightweight, lane-segmentation-driven LDW, which includes a drivable-area fallback for degraded markings, with a two-stage TSD/TSR pipeline consisting of efficient sign localization followed by fine-grained classification, designed to preserve accuracy at embedded latency. To support region-specific deployment under data scarcity, we curate a 35-class local traffic-sign dataset and introduce a generalized augmentation and balancing profile, expanding the corpus from 359 images to 8,750 samples. On an NVIDIA Jetson Orin Nano, the integrated system sustains 15 FPS end to end (32 FPS for LDW and 18 FPS for TSD/TSR). The TSD/TSR branch achieves 97.4 mAP for detection and 95.6% validation accuracy (93.1% on a local real-world set) while maintaining an approximately 13.9 M parameter footprint. Relative to recent methods that target a single task or report results only on desktop-class GPUs, our approach delivers both LDW and TSD/TSR with real-time performance on edge hardware, providing a practical pathway to low-cost, field-ready ADAS in resource-constrained settings.
Figures & tables
Figure 1: Overall Framework of the Proposed Intelligent Driving Assistance System. YOLOv8s [ 31 ] for traffic sign detection, MobileNetV3-small [ 32 ] for traffic sign recognition and TwinLiteNet [ 4 ] for lane detection
Figure 2: Traffic Sign Detection and Recognition Flow
Model
Params
Framework
Precision
Recall
mAP50
mAP50-95
FPS
FR-CNN
42M
PyTorch
-
-
0.595
0.324
0.30
ONNX
0.59
YOLOv5s
7.2M
PyTorch
0.8977
0.7970
0.8665
0.5958
5.40
YOLOv8n
3.2M
PyTorch
0.9826
0.8111
0.9025
0.7192
7.50
YOLOv8s
11.2M
PyTorch
0.9740
0.9092
0.9707
0.7996
3.63
YOLOv8m
25.9M
PyTorch
0.9893
0.8935
0.9540
0.7964
1.32
Table 1: Comparison of Different Models for Traffic Sign Detection (Tested on Dell Latitude, Core i5 8th Gen with 16 GB RAM and 3.6 GHz processor without GPU).
Figure 4: Labels correlogram showing spatial distributions of traffic sign annotations. The concentration of x and y coordinates at the center reflects the alignment achieved after targeted cropping and augmentation.
Parameter
Value
Model Architecture
MobileNetV3-small
Parameters
2.54 Million
Dataset Size
3764 Images
Dataset Split
65% Training, 15% Validation, 20% Testing
Learning Rate
0.0005
Batch Size
32
Table 2: Hyperparameters for MobileNetV3-small
Parameter
Value
Model Accuracy on Validation Set
94.4%
Model Accuracy on Test Set
92.6%
No. of Epochs in Training
46
No. of Images in Inference Set (Self Collected)
58
Correctly Classified
34
Misclassified
24
Table 3: Training Results for MobileNetV3-small
Figure 5: (a) Training and Validation Accuracy (b) Training and Validation Loss
Parameter
Value
Model Architecture
MobileNetV3-small
Parameters
2.54 Million
Dataset Size
8750 Images
Dataset Split
70% Training, 20% Validation, 10% Testing
Learning Rate
0.0005
Batch Size
32
Table 4: Hyperparameters of 2nd Stage Training for MobileNetV3-small
Parameter
Value
Model Accuracy on Validation Set
95.6%
Model Accuracy on Test Set
97.83%
No. of Epochs in Training
18
No. of Images in Inference Set (Self Collected)
58
Correctly Classified
54
Misclassified
4
Table 5: 2nd Stage Training Results for MobileNetV3-small
Figure 6: (a) 2nd Stage Training and Validation Accuracy (b) 2nd Stage Training and Validation Loss
Figure 7: Confusion Matrix of MobileNetV3-small on Test set
Figure 8: Detecting and Classifying Traffic Signs. The traffic sign detected and classified in image (a) No Parking, (b) Cross Roads, (c) Give way, (d) Cross Roads, (e) Speed Breaker and Pedestrian, (f) Speed Limit (40 Kmph) and Slow, (g) U-Turn and (h) Slow
Model
Year
mAP (%)
Accuracy (%)
Params (M)
FPS
Device
Dataset
Pakistani TSR (LeNet-5) [ 28 ]
2018
–
41.0
0.062
–
–
Pakistani TSR
Pakistani TSR (LeNet-5) [ 30 ]
2021
–
83.18
0.062
–
–
Pakistani TSR
Pakistani TSR (Transfer Learning) [ 29 ]
2022
–
55.0
0.062
–
–
Pakistani TSR
Bangladesh TSR (MobileNetV2) [ 41 ]
2023
–
89.65
3.4
–
–
Bangladesh Dataset
Bangladesh TSR (VGG16) [ 41 ]
2023
–
90.31
138
–
–
Bangladesh Dataset
EMDFNet [ 14 ]
2024
97
–
29.7
16.7
RTX 3090
GTSDB
Table 6: Comparison of Different Models for Traffic Sign Detection and Recognition
Figure 9: Block Diagram of Architecture of TwinLiteNet
Figure 10: (a) Shows the orignal result of TwinLiteNet Model (b) shows the Lane Line Segmentation from TwinLiteNet Model
Figure 11: Lane Departure Warning Algorithm
Figure 12: (a) Resizing Frame to 640x360 (b) Cropping Frame for getting ROI i.e. 640x180. (c) Lane Mask from Model. (d) Canny Edge Detection Output. (e) HoughLine Transform Output (f) Placing Mask on ROI
Figure 13: Lane Departure Warning. Red and Blue points indicating center of Right and Left lane respectively while yellow point indicating Vehicle position which is the bottom middle edge of frame. Left lane center comes in threshold window so the Warning Generated for Left lane.
Scenario
Class
Precision
Recall
F-1
Pred. Frames
GT
Accuracy (%)
Day Time
No Warning
0.83
1.00
0.91
194
161
91.56
Right Warning
0.85
0.24
0.37
13
46
Left Warning
0.96
0.96
0.96
386
386
Scenario Totals:
Total Frames: 593
Correct: 543
Night Time
No Warning
0.76
0.97
0.85
1850
1441
85.86
Right Warning
0.97
0.82
0.89
1189
1414
Table 7: Detailed Performance Metrics for Warning Detection Across Different Environmental Conditions
Figure 14: Evaluation of Lane Departure Algorithm in Different Scenarios i.e. (a) Day (b) Night and (c) Rain
Figure 15: (a) Drivable Area Segmentation (b) Edges of Drivable Area Segmentation (c) Classify Left and Right Edge of Drivable Area
Figure 16: (a) shows the Edges of Drivable Area Segmentation (b) and (c) shows the Lane Approximation using midpoint of Left and Right Edge of Drivable Area
Figure 17: Noise in Drivable Area Segmentation due to which lane approximation becomes unstable
Figure 18: Lane departure warning with traffic sign detection and recognition. (a) normal driving with a "No Parking" sign detected and no lane departure. (b) detection of a "No Parking" sign and warning for being too close to the right lane. (c) a "U-Turn" sign and warning for closeness to the left lane. (d) detection of a "Cross Roads" sign and also indicates a lane departure warning for drifting too close to the right lane
Model
Year
TSD
TSR
Object Det
LD
DAS
LDW
Edge Device
Real-Time (FPS)
Multi-task Object Detection & Lane-Aware Models
YOLOP [ 16 ]
2021
✕
✕
✓
✓
✓
✕
Jetson TX2
23
HybridNet [ 17 ]
2022
✕
✕
✓
✓
✓
✕
✕
✕
YOLOPv2 [ 18 ]
2022
✕
✕
✓
✓
✓
✕
✕
✕
CENTERNETS [ 20 ]
2023
✕
✕
✓
✓
✓
✕
✕
✕
YOLOPv3 [ 23 ]
2024
✕
✕
✓
✓
✓
✕
✕
✕
Table 8: Comparison of different models for lane and traffic sign analysis. Here, ✓ indicates that the model provides the corresponding functionality, whereas X indicates that the functionality is not supported.
Figure 19: Overview of proposed experimental setup and data flow. A camera mounted on the vehical dashboard capture real time road scenes. the captured frames are processed on jetson orin nano to perform traffic sign recognition and lane departure warning. the processd result are trasmitted via remote connection for visualization, while warning information delivered back to the driver. The Jetson Orin Nano is powered using external battery.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Class
Traffic Sign
Class
Traffic Sign
0
Bridge Ahead
18
Slow
1
Cross Roads
19
Speed Breaker Ahead
2
Give Way
20
Speed Limit (20 km/h)
3
Left Bend
21
Speed Limit (25 km/h)
4
No Horns
22
Speed Limit (30 km/h)
5
No Mobile Allowed
23
Speed Limit (40 km/h)
Appendix
Table S1: List of 35 Traffic Sign Classes in the Pakistani Traffic Sign Dataset [ 38 ] (Kaggle)
Figure S1: Experimental setup for real-time testing of the ADAS framework on Jetson Orin Nano. (a) Jetson Orin Nano powered through a portable battery. (b) Camera mounted on a tripod connected to the Orin Nano. (c) Local web server visualization of the real-time stream.
Figure S2: Qualitative real-time visualization of the proposed ADAS system. (a) displays the live camera feed from the test environment, and (b) illustrates the processed frame showing accurate lane detection and identification of the “Speed Breaker Ahead” traffic sign.
Figure S3: Qualitative real-time results of the proposed ADAS framework. (a) shows detection of a “Speed Limit 40” sign with no lane marking warning, indicating that the vehicle is at a safe distance. (b) illustrates the same detected sign when the vehicle moves closer to the lane boundary, triggering a lane departure warning on the frame. (c) presents detection of a “Stop” sign with no warning generated, as the vehicle remains within the safe zone.
Hardware
Task
Framework
Processing Unit
Real-Time FPS
Jetson Orin Nano
Lane
PyTorch
GPU
25.50
ONNX
32.00
TensorRT
32.50
Quantized *
22.0
Sign
ONNX
18.00
Lane+Sign
ONNX
15.00
Appendix
Table S2: Performance of Lane Departure Warning, Traffic Sign Recognition, and Combined Task on different hardware platforms. The module marked with * in the “Framework” column for the Jetson Orin Nano denotes that the model was run on CPU.