Vehicle Detection Research Based On Tracking And Clustering Of Feature Point

Vehicle detection method based on video analysis has become a current hot research. Usually, we use the method based on gray difference between target and background and continuous gray variance when a car moves through the virtual detect line to detect vehicles, which can obtain good results when the cameras are installed at a high height and in the middle road. However, some times, we have to put cameras beside the road at a low height due to limitations from road infrastructure, the traditional vehicle detection method based on video analysis may get wrong results because of occlusions between vehicles.In this paper, the research mainly focuses on vehicle detection when cameras are installed beside the road at a low height, and vehicle detection will be performed through clustering the 3D corrected trajectory.There are mainly three parts in vehicle detection based on clustering of 3D trajectory, which are extraction of stable feature points, tracking of feature points based on Bayesian filtering and vehicle detection based on clustering. We use six points whose pixel coordinate and 3D coordinate are known before to calibrate camera. The 3D coordinate according to its pixel coordinate and projection matrix will be calculated by utilizing pedal information. This article has applied target tracking based on bayesian filtering because of the limitation on template matching tracking. Considering the approximate linear motion of target in observation time in 3D space, image trajectory will be obtained from 3D trajectory by using Kalman filter. We get the 3D trajectory of moving vehicle by mapping image tracking points to 3D space. With the detect line in 3D space and the 3D trajectory whose heights are filtered by median filter, we can calculate the mapping relationship between 3D trajectory and its image trajectory again by the height information through median filtering. We finally get the detection result by clustering of 3D trajectory using variance, texture and maximum distance as the similarity measure condition.We have tested the algorithm on different traffic conditions, which shows that the algorithm can detect vehicle effectively and satisfy the real-time and accuracy requirements of the system.