Lightweight Deeplearning Method for Multi-vehicle Object Recognition

The recognition method based on deep learning has a large amount of calculation for the changes different traffic densities in actual environment. In this paper, an integrated YOLOv4-L is proposed reducing computational complexity YOLOv4. characteristics multi-lane flow with were analyzed statistical data sets, and k-means++ clustering algorithm was used to optimize prior frame parameters improve matching degree between frame. GhostNet replace CSPDarknet53 original network structure YOLOv4 as feature extraction network. depthwise separable convolution module introduced 3×3 common network, reduce model detection speed. further improved both accuracy robustness help comprehensive Mosaic enhancement, rate cosine annealing label smoothing. Experimental results show that, Recognition speed greatly at expense minimal reduction: improvement value 47.81%, 49.15% , 56.06% (FPS), respectively free flow, synchronous blocked reduction 2.21%, 0.67%,, 0.05% mAP, respectively.