Three dimensional information estimation and tracking for moving objects detection using two cameras framework

Abstract

Calibration, matching and tracking are major concerns to obtain 3D information consisting of depth, direction and velocity. In finding depth, camera parameters and matched points are two necessary inputs. Depth, direction and matched points can be achieved accurately if cameras are well calibrated using manual traditional calibration. However, most of the manual traditional calibration methods are inconvenient to use because markers or real size of an object in the real world must be provided or known. Self-calibration can solve the traditional calibration limitation, but not on depth and matched points. Other approaches attempted to match corresponding object using 2D visual information without calibration, but they suffer low matching accuracy under huge perspective distortion. This research focuses on achieving 3D information using self-calibrated tracking system. In this system, matching and tracking are done under self-calibrated condition. There are three contributions introduced in this research to achieve the objectives. Firstly, orientation correction is introduced to obtain better relationship matrices for matching purpose during tracking. Secondly, after having relationship matrices another post-processing method, which is status based matching, is introduced for improving object matching result. This proposed matching algorithm is able to achieve almost 90% of matching rate. Depth is estimated after the status based matching. Thirdly, tracking is done based on x-y coordinates and the estimated depth under self-calibrated condition. Results show that the proposed self-calibrated tracking system successfully differentiates the location of objects even under occlusion in the field of view, and is able to determine the direction and the velocity of multiple moving objects.