MDFD2-DETR: A Real-Time Complex Road Object Detection Model Based on Multi-Domain Feature Decomposition and De-Redundancy

Accurate perception of road object information is critical in autonomous driving systems and traffic flow analysis. This task is challenging due to the variability of complex road backgrounds, diversity in object sizes and shapes, and mutual occlusion among objects. To address these challenges, we introduce an end-to-end, real-time model for complex road object detection, named Multi-Domain Feature Decomposition and De-redundancy Detection Transformer (MDFD2-DETR). To the best of our knowledge, this is the first real-time perception model designed for complex traffic scenarios, based on the DETR architecture. In detail, we analyzed the key reasons for the low accuracy of complex road target perception and proposed a Multi-Domain Feature Decomposition (MDFD) module for reducing feature redundancy across spatial, channel, and frequency domains. Additionally, we introduced a Hybrid Position Encoding (HPE) strategy to achieve more effective positional interactions among features to address this issue. To train and evaluate the proposed algorithm, we built a complex, real-world road perception dataset named SWJTU-100k, containing 100,000 samples. Experimental results demonstrate that MDFD2-DETR surpasses existing real-time methods, including YOLO-like, DETR-like models, with fewer parameters and GFLOPs, reaching 82.6 FPS without the use of CUDA acceleration. Further, it demonstrates a 0.5%-1.7% [email protected] improvement on the PASCAL VOC 2012 dataset, a 1.5%-3.3% [email protected] improvement on the SWJTU-100K dataset, a 0.5%-1.5% [email protected] improvement on the KITTI dataset, and a 0.6%-1.5% [email protected] improvement on the Udacity dataset, thus exhibiting its superior performance. In addition, multi-object tracking algorithms based on MDFD2-DETR achieved superior tracking performance in complex road scenes.