Improved DAB-DETR model for irregular traffic obstacles detection in vision based driving environment perception scenario

Machine vision based irregular traffic obstacles recognition plays a pivotal role in the autonomous driving and Advanced Driver Assistance Systems (ADAS) by providing the necessary environment perception capabilities. Traditional models for recognizing irregular traffic obstacles suffer from challenges with small target detection, poor performance in diverse environmental conditions and computational complexity. This work addresses the critical issue of recognizing irregular traffic obstacles in roadway environments. We present an enhanced target detection model based on the Dynamic Anchor Boxes-recognition Transformer (DAB-DETR). The original model’s structure was limited in expressing relative positional information between features due to the reliance on absolute position encoding. To overcome this limitation, the improved DAB-DETR incorporates relative position encoding within the multi-headed self-attention mechanism of the Transformer encoder. Additionally, we propose a novel Average Precision (AP) loss function that unifies classification and localization losses into a single parameterized formula, addressing performance degradation observed in the original model. Experimental results demonstrate significant improvements in detection accuracy for irregular traffic objects, showcasing the effectiveness of the proposed enhancements. According to the testing results, the improved DAB-DETR model’s detection accuracy is 82.00% with Intersection over Union (IoU) equals to 0.5, which is 3.3% better than the original model and 6.20% and 7.71% better than the conventional models, YOLOv5 and Faster R-CNN, respectively.