Research on Real-time Object Detection Algorithm in Traffic Monitoring Scene

Abstract

Aiming at the detection precision and detection speed of Mobilenetv2-YOLOv3 in the process of object detection, the feature extraction module is improved to improve the network feature extraction capability; the FPN multi-feature fusion module and multi-scale aggregation module are used to enhance information between multi-scale feature maps Fusion; Introduce the dilated convolution to build the receptive field module, Improve the ability to extract features of different scale object, and improve detection precision; According to the characteristics of the KITTI data set, K-means algorithm is adopted for dimension clustering to obtain the new anchor box parameter values. Test the algorithm performance on the KITTI dataset, Experimental results show that compared with Mobilenetv2-YOLOv3, the improved algorithm improves the mAP by 8.99%, and the detection speed reaches 14FPS on the embedded hardware TX2 development board.