Vehicle Detection And Tracking Based On Multi-Sensor Fusion

Driving safety is an important problem in traffic, especially in the crowded and complex city road environment. Vehicle detection and tracking become necessary parts of driver assistant system to improve convenience and safety of on-road driving.As for the road situation that exist many obstacles and quick turns, a multi-sensor fusion based vehicle detection and tracking system is proposed. It adopts camera and laser scanner as primary sensors, uses distance information from laser scanner and image information from camera to extract target features which would be utilized for judging vehicle existence. At the same time, it employs Kalman filter and particle filter fusion as tracking context to solve the vehicle dismiss in tracking process happened to quick turns. Experiments show that this system could detect vehicle effectively and realize vehicle tracking correctly and robustly.The architecture that the system relies on is firstly introduced. To solve multi sensors'asynchronous and uncalibrated data, the space co-calibration work based on sensor's model and time calibration work between sensors are done. Calibration is the base of multi-sensor data fusion.Vehicle detection is the base of vehicle tracking. It uses multi-sensor fusion based vehicle detection methods to solve the problem of single sensor's low reliability. Firstly, the regions of interest are extracted through laser data; Then shadow,symmetry and texture are extracted through YUV image,edge image and gray image in turn. Finally the regions that could contain vehicle are marked. Practical experiments show that this method could improve correctness of vehicle detection, lowering the wrong and missing rate.The aim of vehicle tracking is acquiring preceding vehicle's information steadily, then conducting the decision. In city environment, the obstacle disturbance of pedestrians and vehicles and quick turns both effect the vehicle tracking effect, increasing the missing rate. To solve the problem, based on vehicle detection method, a new tracking context combining Kalman filter and particle filter is adopted, and a data association algorithm based on multi-dimension space's Mahalanobis distance is proposed. The tracking system could effectively predict and update vehicle's state using frame information. Experiments under several conditions prove that the system could track preceding vehicle more steadily and correctly, with no missing happened to quick turns and obstacle disturbance.