Aiming at the difficulty of object detection in complex and diverse traffic scenes, including occlusion, we propose an improved CenterNet object detection algorithm, and we optimize CenterNet from four aspects. First, we introduce RFB (Receptive Field Block) to extract and fuse feature information more efficiently Second, we compared different attention mechanisms and selected the attention mechanism suitable for CenterNet object detection, which improved the network's ability to focus on objects. Third, a new method of calculating predicted boxes and ground truth boxes is adopted, which further accelerates the training convergence speed of the network. Finally, in view of the problems of single lighting and fixed position in the data set, we introduced a data enhancement method to effectively improve the robustness of the network. After experimental verification, the average accuracy of our improved network on the PASCAL VOC dataset has increased by 4.340/0, and in the experiment of re-detection of traffic scenes, the accuracy has increased by 3 %, reaching 89%. Compared with the original CenterNet, the improved CenterNet is more suitable for object detection applications in traffic scenes.
Improvement Based on CenterNet and Application of Traffic Scenarios
2023-10-28
1826561 byte
Conference paper
Electronic Resource
English
Ear Detection based on CenterNet
IEEE | 2020
|Vehicle Light Detection Method Based on Improved CenterNet
Springer Verlag | 2022
|