Self-adaptive ground calibration in binocular surveillance system

Object detection and tracking have always been crucial and challenging topics in computer vision. Compared with monocular vision systems, binocular vision systems (BVSs) have the advantage of dealing with illumination variation, shadow interference, and severe occlusion. Usually, the BVS constructs the world coordinates system by manually calibrating the ground plane. However, the camera vibrations decreases the calibration precision and weakens the system performance. To automatically correct and update the parameters of ground plane, we introduce Linear Discriminant Analysis (LDA) method to analyze the results of object localization and include the feedback in the surveillance system, in this way, a close loop system that greatly improves the accuracy and stability of surveillance system is constructed. Experimental results demonstrate that our approach works well in BVS for video surveillance.