Fast Vehicle Detection Algorithm Based On Deep Learning

It is one of the important tasks of the road monitoring system to detect and count the number of vehicles on the road in real time.When the traffic on the road is large,the vehicles block each other,so the image detection algorithm is difficult to detect the vehicles in the road monitoring video in real time and accurately.In order to improve the detection rate of occluded vehicles,this paper proposes to use windows instead of vehicle body as the target.By using residual connection and multi-scale feature extraction,a detection network with only 24 convolutions is constructed.This method can effectively improve the detection rate of mutual occlusion vehicles,and can detect vehicles in high frame rate surveillance video in real time.The experimental results show that the recall rate of the algorithm is 95%,the accuracy rate is 96%,and the detection speed in GeForce GTX 1080 Ti video card environment is 142 FPS.In order to make the detection algorithm proposed in this paper have better effect on the detection of complex objects,it is necessary to increase the depth of neural network to extract enough feature information.Therefore,based on the feature extraction network of darknet53,this paper proposes to add a multi-scale void convolution module to the detection network.Thus,the association between local features and the whole feature map can be increased and the feature information can be strengthened,so that the accuracy of vehicle detection based on window features can be further improved in complex cases.In order to make the detection algorithm can be effectively arranged to the edge,it is necessary to further reduce the memory consumption of the weight model and speed up the detection speed of the algorithm.Therefore,this paper proposes a model channel pruning strategy to compress the model.For each channel,a scaling factor γ is introduced to multiply the output of the channel,and the network weight and these scaling factors are jointly trained.Finally,the channel with small scale factor is removed and the pruned network is fine tuned.Thus,the final generation model improves the detection speed and reduces the memory consumption under the condition of similar detection accuracy,and achieves better results.