Research On Vehicle Detection Algorithm Based On Inverse Projection Transformation

Vehicle detection technology based on video is a branch of intelligent transportation system,it can be divided into the method based on appearance characteristics and motion feature.Though some advances have been made,many problems exist,such as vehicle occlusion and adhesion,camera image distortion and so on.In view of the above problems,this paper proposes a vehicle detection algorithm based on inverse projection transformation.Firstly,according to the principle of camera imaging,projection transformation of 3D scene and 2D image is established,calibration of traffic scene is completed,the inverse projection plane is designed,the information on the inverse projection plane is reconstructed,to obtain inverse projection map.The vehicle target can be replaced by detecting organic combination of vehicle parts or amalgamation of multiple features.At night,the headlights selected as detection object,headlights are first detected according to the geometric characteristics,Gauss Mixture Model is established through distance between the headlights and height of the lights,this probability model is used to complete the accurate detection.Problem of improper matching exists,then positive and negative samples of LBP(Local Binary Patterns)features are extracted from inverse projection map,using these LBP features to train AdaBoost classifier to complete detection,this method can effectively solve the problem of improper matching.By day,taillights and license plate are selected to detect as the object,through color model to detect taillights and license plate,and then through the geometric relationship between the parts to establish a Markov Random Field to complete vehicle detection.This method depends on the color information of red taillights and blue license plate,red and blue car cannot be detected.The features of vehicle symmetry fixed in the inverse projection map and vehicle shadow are selected to detect,first use Symmetrical SURF algorithm to detect the symmetry characteristics of the target,then to determine whether the targets are vehicles according to the characteristic of vehicle shadow.This method can effectively solve the problem of missed detection of red or blue car.In this paper,the algorithms is compared with actual traffic scenes,experimental results show the algorithms can achieve good effect on the vehicle detection and solve some problems.At night the detection accuracy is increased from 93.5% to 96.5%.By day detection accuracy is 93.8% by detecting taillights and license plate,and detection accuracy is 95.5% by detecting vehicle shadow.