Lightweight Deeplearning Method for Multi-vehicle Object Recognition

Abstract

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