Self-adaptive Ground Calibration In Binocular Vision System For Object Detection And Tracking

As one of the most vital research subjects in computer vision,object detection and tracking has become the key technology by now.Being widely applied into different applications,it becomes a part of our life,such as human-computer interaction,public safety detection and military applications.Currently,there is still much room for its improvement.This paper aims to improve the performance of object detection and tracking,thus to enhance the tracking accuracy and stability.Firstly,for monocular camera method,it's hard to handle problems such as multi object occlusion,shadow interference,light changes,and various appearance of objects.Based on binocular vision system,we reconstruct 3D information of objects to overcome object occlusion and shadow interference,thus to obtain better tracking results.Secondly,many tracking methods based on stereo vision obtained the dense depth map to construct 3D model of real objects.But it's time-consuming for calculating depth information of each pixel.By setting tracking region to filter out background feature points,we extract sparse feature points of objects for object detection,so as to accelerate tracking process.According to the characteristics of the object shape,kernel density estimation is used to model probability space for feature points.Feature points belong to one object are clustered into one group by mean-shift method.For pedestrian,projection on the ground is similar to ellipse,so constructing elliptic kernel function when clustering will accelerate detecting process as well as avoiding from interference of various appearance.Then,Kalman filter algorithm is use to tracking objects by detection results.Then,we constructing world coordinates system by calibrating the ground parameters,but once the cameras shift after ground calibration process,tracking results will be serious influenced.For upright and slender objects,we find a plane that separates feature points of different object clearly by LDA algorithm.Then,we correct the calibrated ground parameters automatically,thus to enhance the system stability.Finally,large amount of experiments has been conducted to show that clustering for sparse feature points will performs well before tracking by Kalman filter method.Besides,LDA algorithm will improve the stability of the system effectively.