ADVANCED REAL-TIME TRAFFIC SIGN SEGMENTATION AND CLASSIFICATION USING HYBRID DEEP CONVOLUTIONAL ARCHITECTURES ON GTSDB AND BENCHMARK DATASETS
TL;DR
A rigorous comparative analysis of deep learning models, specifically U-Net, Mask R-CNN, and YOLOv8-Seg, for real-time traffic sign segmentation is presented, finding that YOLOv8-Seg achieves an optimal trade-off with a mean Average Precision.
Abstract
Autonomous driving systems and Advanced Driver Assistance Systems (ADAS) heavily rely on precise visual recognition of traffic control infrastructure under varying real-world conditions. While traditional object detection methods utilize bounding boxes, pixel-level semantic and instance segmentation provide essential geometric detail required for motion planning and vehicle control. This paper presents a rigorous comparative analysis of deep learning models, specifically U-Net, Mask R-CNN, and YOLOv8-Seg, for real-time traffic sign segmentation. Training and evaluation were conducted on benchmark datasets including the German Traffic Sign Detection Benchmark (GTSDB) and TT100K. To address extreme pixel-level class imbalance inherent in roadside imagery, a hybrid compound loss function combining Focal Loss and Dice Loss was developed. Experimental results indicate that YOLOv8-Seg achieves an optimal trade-off with a mean Average Precision (mAP50) of 89.7% and Intersection over Union (IoU) of 81.3% while operating at 45.2 Frames Per Second (FPS) on standard embedded GPU hardware. U-Net demonstrated superior boundary reconstruction accuracy with an IoU of 82.1%, whereas Mask R-CNN achieved peak precision (91.2% mAP50) at reduced inference speeds (12.4 FPS). The source code implementation and optimized hyperparameter setup in PyTorch 2.0 are detailed.