Research On Vehicle Detection And Tracking Method Based On Machine Learning

Nowadays,with the rapid development of economics and economic globalization,the development of automobiles in China has given people great convenience and improved people's life quality.However,with the increasing number of vehicles,the road conditions in China are not optimistic.The advent of Advanced Driver Assistance System(ADAS)can effectively prevent dangerous driving.The important part of ADAS is the collision avoidance module and lane departure warning module.In the system,the function of the collision avoidance module is that the forward vehicles can be monitored in real time,and detected and tracked.Using the designed vehicle detection and tracking algorithm,the system can access to the location information of the forward vehicle,and on the basis of this,the distance and direction between the vehicles can be determined.For the forward vehicle collision avoidance system,the key point lies in the forward vehicle detection and tracking algorithm.The research of detection and tracking of forward vehicle is studied in this paper including the following:(1)The second chapter introduces the mainstream vehicle detection and tracking methods based on machine learning in recent years,and explains the advantages and disadvantages of each method,which lays the foundation for follow-up vehicle detection and tracking algorithm research.(2)In the third chapter,we study the forward vehicle detection algorithm based on machine learning.The main research work includes:constructing the positive and negative sample sets of training detectors;designing the classifier which consists of a number of weak classifiers to form a strong classifier,and each weak classifier corresponds to a feature,the AdaBoost algorithm is used to select the weak classifier which has good performance for vehicle detection;considering the vehicle detection process,most of the region in the image does not contain the vehicle detection object,so the cascade strong classifier is suitable for such applications.In the experiment,LBP,Haar and HOG are used for training,and a cascade classifier with multiple layers is designed.Experiments show that the proposed method can effectively detect the forward vehicle in different environment and has good robustness to meet the real-time requirements of vehicle detection.(3)In the fourth chapter,we study several target tracking algorithms and propose an improved KCF forward vehicle tracking algorithm.The traditional KCF tracking algorithm has the problems of tracking scale and target loss.To solve this problem,this paper presents a method of combining the scale filter of the correlation filter algorithm to improve the KCF tracking algorithm.Firstly,the target vehicle is obtained by the position classifier in the traditional KCF algorithm,the scale pyramid is established for the current target vehicle,and the maximum scale response is obtained by the scale-dependent filter to obtain the current target scale information,and finally the new target image is used as a training sample to update the target position model and scale model.The experimental results show that the improved KCF algorithm in this paper has a good effect on the tracking of vehicles in front,and has good robustness and tracking accuracy in a variety of environments and vehicle changes.