Research On Vehicle Detection Algorithm Based On Convolutional Neural Network

In recent years,computer vision has become a research hotspot in the field of intelligent transportation due to the great improvement of computer computing power.It has received extensive attention from Chinese and foreign scholars and has achieved certain research results.However,there are still many difficulties in the actual application process that need to be solved.The problem mainly focuses on two aspects: First,due to the limitation of computer hardware computing power,when computer vision technology is applied in the field of traffic supervision,the real-time detection,the algorithm itself can not be too complicated,so this has a certain degree of contradictions to accuracy of detection.Secondly,the traffic scene is different from the general detection scene.There are many kinds of objects in the scene,and the factors such as illumination,shadow and occlusion seriously affect the recognition accuracy and positioning accuracy.How to design the algorithm to reduce the influence of these factors without affecting the real-time performance,and improve the robustness of the detection algorithm has become the focus of researchers.This paper reviews the relevant literature and research status in the field of target detection,and compares and summarizes the existing research results.On this basis,the research objectives and research contents of this paper are determined.This paper is dedicated to the realization of vehicle detection from the perspective of road driving through self-built data sets.The research content is mainly divided into three parts.Firstly,the related concepts of convolutional neural networks are used as the entry point.The basic ideas,components and parameter training principles of convolutional neural networks are elaborated,which lays the theoretical foundation for the YOLOv3 model and Faster R-CNN model.Secondly,based on the theory of the previous chapter,the road vehicle rapid detection under the YOLOv3 model is realized.The k-means algorithm is used to cluster and analyze the scale and aspect ratio of the real bounding box of the vehicle samples in the self-built data set.9 real bounding boxes for the vehicle sample.Then the clustering effect of k-means algorithm has dependence on the initial clustering center.The k-means++ algorithm is used to re-cluster the vehicle samples in the database to obtain the optimized a priori bounding box.Experiments show that the bounding box optimized by the means++ algorithm can improve the detection accuracy.Finally,based on the cluster analysis of the real bounding box scale and aspect ratio of the vehicle samples in the self-built data set,the priori box scale under the Faster R-CNN model is rationally optimized.Accurate detection of road vehicles under road scenes is achieved.Combined with experimental results verification and analysis.After the prior frame optimization,the detection accuracy is partially improved.