The Research Of Forward Vehicle Detection And Tracking Based On Convolutional Neural Network

With the development of domestic road construction and the increasing of vehicle ownership,the number of road traffic accidents,especially the vehicle forward collision accidents is rising in high speed.When the risk arises,Forward collision warning system can effectively prevent the forward collision accidents,and the vehicle detection technology is a key part of the system.This paper analyzes the existing technologies related to vehicle detection,focusing on forward vehicle detection and tracking method in the daytime.The main work includes:(1)Vehicle hypothesis generation based on the shadow and width constraint of vehicleFirstly generate shadow regions by the method of HVAM;Secondly,do morphological processing to the shadow binary image generated in last step;Thirdly,eliminate a large number of shadows with no vehicles by constraints of vehicle attributes.In the end,For there may be situation that some vehicle hypothesis are overlapping,we analyzed many situations,and eliminated the overlapping hypothesis,then we got the regions of vehicle hypothesis.(2)Vehicle hypothesis validation based on CNNFirstly,we built a database of vehicle;Secondly,we designed and trained several different structures of CNN net;Lastly,aiming at GTI database,we compared and analyzed the validation results of traditional methods and CNN method,coming to a conclusion that using CNN method to detect vehicles is superior to traditional algorithms.(3)Vehicle tracking based on CNNWe proposed a vehicle kalman filter tracking method based on CNN.At first,we used kalman filter algorithm for state prediction,and then we observed the adjacent area of targets using CNN,completing the process of tracking;however,kalman filtering algorithm has a limitation that it can only process targets one by one,thus,it consuming much time.Aiming at the problem,we further proposed a regional tracking method based on CNN.By experimental verification,the proposed regional tracking method can greatly improve the real-time performance,ensuring the accuracy and stability in the meantime.