End-to-End Object Detection by Sparse R-CNN With Hybrid Matching in Complex Traffic Scenes

Object detection is an important research topic in traffic scenes. Although object detection based on deep learning has made some progress, it is still a challenge to improve the accuracy of multiple objects in complex traffic scenes. One-to-one matching is the key for Sparse R-CNN to achieve end-to-end object detection, enables detection without non-maximum suppression (NMS) to remove redundant boxes. However, the matching strategy makes the number of learnable proposal boxes assigned as positive samples low, and significantly reduces the training efficiency of positive samples. Therefore, we use a hybrid matching strategy, which combines the original one-to-one matching branch with a new one-to-many matching branch during training to improve the accuracy of detection, and uses the original one-to-one matching during inference, thus maintaining the end-to-end advantages and the same inference efficiency. Meanwhile, we propose the gated channel space attention (GCSA) to build more powerful backbone extraction features, and dynamic space multi-scale fusion (DSMSF) to fuse multi-scale and multi-location spatial information to obtain richer global features. Due to the complexity of the traffic scenes, we not only validate the performance of the model with the KITTI dataset and Mini-COCO-Traffic dataset, but also demonstrate the generalization ability of the model with the COCO 17 dataset and the VOC dataset. The experimental results show that our model improves the accuracy significantly on all four datasets, including 2.4%
AP
and 2.6%
AP.S
with a model complexity of 137.8
GFLOPs
on the KITTI dataset, and a object detection model with a Resnet50 backbone achieves 62.4%
AP@0.5
on the COCO dataset.