Detection Of Vehicle Parts Based On Convolution Neural Network

With the development of the times and the progress of the science and technology,the number of cars has increased exponentially,meanwhile,traffic congestion,environmental pollution,frequent traffic accidents and other issues are more and more.In this case,Intelligent transport system came into being.As an important application of Intelligent transport system,vehicle parts detection can effectively solve the vehicle theft,traffic accident escape and other vehicle crime problem.Vision-Based detection of vehicle parts be a commonly used method with the low cost of video equipment and no position requirements of laying and other advantages.This paper presents a vehicle parts detection algorithm based on convolution neural network for images.The algorithm uses the depth convolution neural network(ZF network)to extract the local features of the image,avoiding the process of man-defined features.The use of region proposal network instead of selective search greatly saving the time of generate candidate region.In the use of region proposal network,this paper uses the average aspect ratio of the vehicle components and the statistical information of the area to set the aspect ratio and the scaling factor of the initial area of the network input to meet the practical application requirements.Considering the vehicle as a whole,in this paper,a classification and adjustment method based on the absolute position distribution constraint of each component is proposed.On the basis of this,an improved method which use the relative position Gaussian model has been proposed.The method utilizes the spatial position prior information of the vehicle parts to improve the performance of the vehicle parts detection algorithm.Finally,an angle and distance network used to extract the global features of the region to regression the bounding box and classification with the local features has been proposed.The method greatly improves the accuracy of the vehicle parts detection algorithm.The network makes full use of the contextual relationship between vehicle components,and obtains a fast and accurate vehicle component detection algorithm.