| The one of key problems for autonomous mobile robot is human target detection and tracking, in which foreground measurement is a kernel. Traditional tracking methods using a monocular vision search a target in whole image on the base of the known model. These methods are usually limited to the known model, need high computation cost and do not have the real-time performance. In real applications, existing methods are often affected by some factors, for example, target randomly appearing or disappearing, movement multiplicity, occlusion and rotation of a target.In order to improve the accuracy of human target detection, this paper synthesizes the characteristics of laser and camera system to construct an extended 2D environmental paradigm. A new human model is designed using the above environmental paradigm, which includes depth, height and color information and favors to detect a human target.This paper realizes multi-human targets detection and tracking in a Bayesian framework. In the process of target detections, candidates are firstly extracted by a laser sensor, and then the regions of target in image are acquired according to coordinate transformation in extended 2D environment. Finally human targets are identified by multipart detection strategy. In addition, multipart distributed color body model and dynamically updated strategy are introduced to eliminate the effects of lighting change, target rotation and partial occlusion. Furthermore, a variable velocity model is proposed to forecast the position of a target. Experiment results with the mobile robot and data analysis show the method's validity and practicability.Furthermore, a human target tracking method is presented in multi-robot cooperation domain. The method can track human targets in a cluttered scenario by a "divide and conquer" strategy and make up for the deficiency of single robot system. The experiment on a real platform validates the efficiency and robustness of the proposed method. |
PDF Full Text Request