HRT-DETR: efficient traffic sign detection
TL;DR
Experiments show that HRT-DETR outperforms RT-DETR and other mainstream detectors in both accuracy and inference speed, demonstrating its effectiveness in dense small-object traffic sign detection under complex backgrounds and its superior real-time performance.
Abstract
To balance accuracy and speed in traffic sign detection, this paper presents HRT-DETR, an enhanced end-to-end detector built upon RT-DETR. Although RT-DETR eliminates the need for Non-Maximum Suppression (NMS) and enables real-time inference, its hybrid encoder discards fine-grained spatial details, resulting in suboptimal performance for dense small-object detection in traffic sign scenarios. Its query selection mechanism is also highly sensitive to minor deviations in initial bounding boxes, limiting robustness under scale variations.To overcome these limitations, HRTDETR introduces three key improvements:(1) A High-efficiency Local Attention (HLA) module is added to the backbone to enhance small-target feature representation via pyramid and bidirectional pooling while maintaining low computational cost;(2) The encoder is redesigned as a two-layer dilated structure, where the first layer extracts multiscale contextual cues and the second employs AIFI for cross-scale semantic fusion, enriching feature representation and improving adaptability to complex scenes;(3) A Quick Intersection Over Union computation module (QIOU) is integrated between the encoder and decoder to accelerate IOU-based quality assessment through precomputation and parallelization, improving query initialization and localization reliability.Experiments on the CCTSDB 2021 benchmark show that HRT-DETR outperforms RT-DETR and other mainstream detectors in both accuracy and inference speed, demonstrating its effectiveness in dense small-object traffic sign detection under complex backgrounds and its superior real-time performance.