Research on multi-target detection in complex traffic environment based on improved DINO

To address the issue of low detection accuracy for small and occluded objects in complex scenarios for autonomous driving perception tasks, we propose an improved multi-object detection algorithm based on end-to-end object detection with transformers with Improved deNoising anchOr box (DINO). The bidirectional feature pyramid Network (BiFPN) structure is incorporated into our approach to further fuse high- and low-level features, thereby enhancing the extraction capability for multi-scale features. Experiments conducted on the BDD100K and SODA10M datasets demonstrate that the improved model achieved an increase of 2.4% and 2.3% in the mean average precision for vehicles and trucks, respectively, and an increase of 5.4% and 4.4% in the accuracy for pedestrians and cyclists. In addition, the improved model was converted to Open Neural Network Exchange Intermediate Representation format to test its generalization capabilities in various campus environments.