Research On Vehicle Detection Algorithm Based On Lightweight Network

With the rapid growth of the economy,the occupancy ratio of vehicles in China has increased year by year.Road congestion,frequent accidents,and traffic safety problems have become increasingly prominent.To ensure driving safety and strengthen vehicle management,safe and efficient vehicle detection technology is urgent needs.Vehicle detection based on deep learning plays a vital role in various fields.It presents a major development direction to computer vision in recent years.Research on vehicle detection algorithm based on lightweight network is an important research topic.Lightweight vehicle detection includes the exploration of network structure and computing efficiency,and is widely used in many fields such as intelligent transportation,and so on.However,there are many problems in traffic scenarios,such as vehicle targets scale changes in the camera,vehicles overlap each other,and the model lacks of feature extraction ability.These situations affect the accuracy of the network’s judgment of various categories of vehicles,and pose a threat to urban traffic safety.Therefore,it is meaningful to solve the above problems in lightweight network vehicle detection.Based on these problems,this thesis proposes two solutions.An improved vehicle detection algorithm based on the lightweight network Yolox-nano is proposed to enhance the feature extraction ability of the lightweight network.Firstly,the feature extraction ability of the network is enhanced by Ghost module.Ghost module generates part of the vanilla feature maps by using fewer convolution kernels,and then generates cheap feature maps efficiently by cheap operation.Ghost module can control the ratio of the feature maps generated by vanilla convolution and cheap operation through variables.Then,by reconstructing the activation function inside the Ghost module,the nonlinear ability of the network is enhanced.Finally,the convolution layer inside the Ghost module and the batch normalization are fused to accelerate the network inference time.Finally,experiments show that the improved network can increase 4.9% and 3.1% m AP on Pascal VOC and MS COCO vehicle datasets respectively,and the improved network is 11.4% and5.3% m AP higher than Yolov4-tiny.The feature extraction ability of the improved network is better,and the detection effect is remarkable compared with the original network.A lightweight vehicle detection algorithm based on long-distance dependence and multi-scale expression is proposed to enhance the anti occlusion ability and multi-scale representation ability of lightweight networks Yolov5 s.First,the long-distance dependence is captured by the Visual Attention Network,so that the network can perceive the relationship between feature points.Secondly,the horizontal residual is constructed again in the residual structure,and the same number and different sizes of receptive fields features are constructed in a residual structure to enhance the multi-scale representation ability of the network.Finally,experiments show that the improved network can increase 2.1% and 1.7% m AP on Pascal VOC and MS COCO vehicle datasets respectively,and the improved network is 13.9% and9.7% m AP higher than Yolov4-tiny.The anti occlusion ability and multi-scale representation ability of the network are stronger than before,and the detection results are more competitive than the original network.