Traffic Sign Recognition for Autonomous Driving Using Branched YOLOv2 with Geometric Verification

Forum topic · 小凯 · 2026-09-22

Summary

This paper presents a traffic sign recognition (TSR) system for autonomous driving based on YOLOv2, performing detection and classification simultaneously. Two complementary improvements are explored: (1) a branched YOLOv2 architecture with intermediate prediction layers that terminates inference early for easy cases, using both whole-image and cell-wise branching strategies; and (2) geometric information to reduce confusion between visually similar signs, obtained via unsupervised Bayesian image segmentation and compared against class-specific geometric templates inside YOLOv2's bounding boxes, usable at inference or as an auxiliary training signal. A dedicated dataset was built by combining GTSDB and GTSRB samples through seamless cloning and controlled image transformations, covering ten sign classes with 3,000 training and 300 test samples. The selected branched model runs in 0.647 seconds with 0.680 mAP versus baseline YOLOv2 at 0.6607 seconds and 0.680 mAP; adding geometric validation at inference raises mAP to 0.713, while the geometry-trained variant reaches 0.697 mAP at 0.6608 seconds. Paper by Arefeh Rezaei, arXiv:2609.22060.

Paper Overview

Field: Computer Vision Author: Arefeh Rezaei Date: 2026-09-18 arXiv: 2609.22060

Abstract

Traffic sign recognition (TSR) is an important perception task for autonomous driving and advanced driver-assistance systems, where a system must both localize traffic signs and determine their semantic classes efficiently. This work presents a TSR system based on YOLOv2 for simultaneous detection and classification. Two complementary modifications are studied. First, YOLOv2 is extended with intermediate prediction layers, forming a branched architecture that can terminate inference early for easy cases and reduce computation time. Both whole-image and cell-wise branching strategies are investigated. Second, geometric information is introduced to reduce classification errors between visually similar signs. An unsupervised Bayesian image-segmentation method produces binary representations that are compared with class-specific geometric templates within YOLOv2's bounding boxes; this information can be used either at inference time or as an additional signal during training.

Dataset and Results

A dedicated dataset was constructed by combining GTSDB and GTSRB samples using seamless cloning and controlled image transformations.
Experiments cover ten traffic sign classes: 3,000 training samples and 300 test samples.
Selected branched architecture: runtime 0.647 s, mAP 0.680 (baseline YOLOv2: 0.6607 s, mAP 0.680).
Geometric validation at inference: mAP 0.713.
Geometry-trained variant: mAP 0.697, runtime 0.6608 s.