Real‐Time Tire Tread Identification From Wheel Tracks Using YOLOv11 and Triplet Embedding

This paper presents a unified real‐time framework for tire tread identification that integrates tire‐mark detection and tread retrieval into a single end‐to‐end pipeline. Unlike existing studies that directly apply standard YOLO or metric‐learning models, the proposed system introduces three key innovations: (1) a YOLOv11‐based detector customized with adaptive anchor re‐estimation and balanced focal loss, which significantly improves localization on blurred, partial, or low‐contrast tire impressions commonly found in forensic scenes; (2) a ConvNeXt‐ECA embedding head within a triplet network, designed to capture fine‐grained ridge textures under varying illumination and heterogeneous surface conditions; and (3) the construction of one of the first diverse forensic tire‐mark datasets in Vietnam, supporting improved robustness across diverse scenes and benchmark evaluation for tread retrieval. Experimental results show that the enhanced YOLOv11 detector achieves 96.52% precision and 93.91% recall, while the proposed retrieval module obtains a Top‐1 accuracy of 84.78% and mAP of 91.28%. Together, these contributions support practical forensic decision‐making in real‐world scenarios.