Research On Vehicle Detection Method Based On Deep Learning

The detection and identification of road vehicles is one of the key technologies of intelligent transportation systems.As the number of vehicles in my country continues to grow,the problems of traffic accidents and road congestion become more and more serious.For the traffic supervision department,accurately grasping the traffic volume of traffic is conducive to timely traffic unblocking and reducing the probability of various traffic accidents.Therefore,real-time positioning and identification of road vehicles is of great significance.At present,camera equipment is widely installed on roads.Image-based vehicle detection and recognition does not require excessive hardware cost for road equipment,and the cost performance is very high.The thesis detects and recognizes road vehicles based on images,and selects deep convolutional neural networks to learn and predict vehicle images.In order to meet the requirements of real-time and accuracy of road vehicle detection and recognition,the paper uses a single-level target detection framework YOLO v3 as the basic model,and proposes an improved model of vehicle target detection and recognition MSCSE-YOLO.The main improvement strategies are divided into the following three points:(1)The convolution kernel in the YOLO v3 backbone network extracts features from the single-channel area of the feature map.This method mainly extracts the height and width information in the area,ignoring the characteristics of the interaction information between the various channels.The paper introduces the Senet feature weight optimization structure into the backbone network of YOLO v3,assists the convolutional layer to learn the feature information between the channels,promotes the expression ability of the feature map with greater effect,and balances the features of the three dimensions of the feature map length,width and height.Information,improve the learning ability of the convolution check on the input image;(2)In the YOLO v3 backbone network,the traditional convolution process generates a huge amount of calculation,and reducing the amount of calculation in this process will increase the detection speed of the model.The paper performs convolution depth separation on the convolutional layer of the backbone network,performs one-to-one convolution on the input feature map with a single-channel convolution kernel,and then uses 1x1 convolution kernel for channel dimension conversion.This method does not affect the input and output feature map.Shape,but the amount of calculation will be greatly reduced;(3)The maximum effective area of the cell of the output feature map of the YOLO v3 backbone network corresponding to the input image is too large,making the foreground target area small in the effective area area,useful feature information Insufficient proportions are not conducive to the expression of foreground vehicle features,nor to the detection of small targets by the model.The paper adds a smaller scale output node to YOLO v3,and adjusts the number of layers of the backbone network to increase the proportion of foreground vehicle areas in the effective area in the smallest scale.The experimental results on the bit-vehicle data set show that the channel weight optimization strategy has little effect on the detection speed,and the detection mAP value has increased by 2.41%;the convolution depth separation strategy detection mAP has decreased by 2.79%,but the detection speed has increased by nearly 91%;small-scale increase strategy detection mAP has increased by 5.64%,detection speed has been reduced by 37%,the MSCSE-YOLO detection mAP value of the three integrated strategies has increased by 6.25%,detection speed has increased by 17%,and the average detection accuracy has been compared with the speed Both have been greatly improved in YOLO v3,which shows the effectiveness of the improved model of the paper on performance improvement.