Research And Implementation Of Hierarchical Refinement Based Vehicle Detection Framework

At present,in order to better manage the worsening traffic situation,more and more cameras are installed on the road to form a large amount of video data.Therefore,it is of great significance to study how to extract vehicle target quickly and accurately from videoThis paper first discusses the basic theory and methods of target detection,as well as the current research situation at home and abroad,and then discusses the latest research results.Among them,the target detection method based on depth learning has achieved a good detection effect,and has gradually become the focus of research in this field.The main work of this thesis focuses on the combination of various vehicle target detection methods:(1)Background difference based target detection algorithm: background subtraction detection algorithm is the earliest target detection algorithm.It has the characteristics of fast speed and low detection rate.The algorithm is simple,efficient and widely used.Almost all of the early target detection algorithms were based on the background subtraction algorithm.But background subtraction algorithm has some shortcomings,for example,it can not effectively overcome the problem of false detection of Shadows.(2)Adaboost target detection algorithm based on Adaboost based target detection algorithm is the target detection algorithm,the algorithm first successfully used in face detection,face detection is the first to do real-time algorithm,and overcome the influence of illumination and posture changes,but the computational complexity of this algorithm relatively high false detection caused more,can not be applied in many embedded devices.(3)Depth image recognition method based on Neural Network: in recent years,the depth of the neural network to set off a revolution of artificial intelligence,the image recognition accuracy to new heights,such as face recognition accuracy rate has more than human.But the computational complexity is so high that GPU is generally needed to achieve real-time performance.(4)The background difference,the target detection algorithm of Adaboost cascade and depth neural network: a lot of experiments show that the algorithm has low computational complexity and poor background,can detect the foreground region efficiently,so the proposed image background difference,eliminate background region,only running the Adaboost target detection algorithm in the foreground region.Finally,according to the Adaboost output target results using two times the depth of the neural network identification,filtering out the vehicle target error detection.In this way,we can make good use of the advantages of the above algorithms,and avoid their respective shortcomings.At the same time,this paper improves the Adaboost target detection algorithm,that is,to reduce the number of Adaboost scanning window by calibrating the target size in different positions of the image.