Research On Vehicle Detection Algorithm Based On Key Point

The vehicle detection model is susceptible to the influence of illumination,occlusion,geometric deformation and other factors,and the vehicle size in the image will change with the distance of the camera lens,resulting in the vehicle detection model to the distant vehicle identification is not high.At present,there are a large number of vehicle detection algorithms based on deep learning,most of which use anchor box mechanism as the pre-processing method of candidate vehicle detection boxes.The anchor box mechanism can improve the detection accuracy and detection time of the model to a certain extent,but its various problems lead to certain limitations of the vehicle detection algorithm.Aiming at the above problems,three kinds of vehicle detection algorithms are proposed in this paper: vehicle detection algorithm based on corner point,vehicle detection algorithm based on center point and vehicle detection algorithm based on key point triplet.The main work is as follows:(1)Vehicle detection algorithm based on corner point.Considering that the mainstream vehicle detection algorithms adopt anchor box mechanism to extract candidate areas,this will cause difficult parameter selection and imbalance of positive and negative samples,this paper uses the corner of the vehicle to replace the anchor box,and detects the vehicle by judging whether every pixel in the image is the corner of the target vehicle.Due to the occlusion and light of vehicle will significantly impact on the precision of the model,using gamma transform and random erasing strategy to enhance vehicle images,in the process of vehicle corner feature extraction,a dense connection block and residual unit for feature extraction network reconstruction,improve the corner feature extraction ability of the network,and to increase attention to small scale information for small target prediction.(2)Vehicle detection algorithm based on center point.There is one step is to match vehicle corner point when detecting vehicle corner point,aim to determine the corner point belonging to the same target vehicle and regression prediction box of the vehicle,and the corner point of vehicle is divided into the upper left corner point and the lower right corner point,it is more complex than detecting center point,this paper adopts the method of testing center of vehicles to prediction the position of the vehicle and categories.In view of the feature extraction networks need to satisfy the input characteristics of scale deposit,and the size of the vehicle images often do not meet this requirement,when making the adjustment of the size and there is the possibility of a vehicle geometry deformation,this paper adopts the method of deformable convolution instead of traditional convolution in order to improve the extraction ability for the geometric deformation feature,GIo U loss item is increased to predict the position deviation between prediction box and the real box.(3)Vehicle detection algorithm fusing corner point and center point.In this paper,the feature information of the corner point and the center point is fused to improve the vehicle detection performance of the model.In order to improve the detection time of the model,this paper uses the fire module to compress the feature extraction network,and uses the dilation convolution to increase the receptive field of the center point feature,which can not only decrease the inference time of the model but also make the model learn more features of the vehicle key points.