Object Tracking In Complex Environment And Its Application In Mobile Robot

Object tracking has attracted more and more research interests in the field of computer vision. It is an omnipresent elementary task in a variety of applications including surveillance, gesture human-machine interface, motion capture, video editing and compressing, visual serving, etc. However difficulties to achieve robust and accurate performance in tracking objects arise from abrupt object movement, changing appearance patterns of the target and the scene, and partial or complete occlusions of the interest entities. It is an active topic in the field of computer vision.Financed by the project of"Intelligent Household Robots for Security and Service", in this thesis we focus on how to achieve robust and accurate performance in tracking objects and apply this method into mobile robot.The main content of this paper is as follows:(1) We propose a new approach based on correlation matrix which integrates multiple features, including both spatial and statistical properties, to describe the interest object. It can separate the object from complex background. Besides, we suggest a method to obtain the best matching region by calculating a centroid estimated from a set of candidate points in searching window. Comparing with covariance matrix based global search method which takes the point with the smallest distance to the target, our method can achieve better performance when tracking object in complex environment when the illumination change can be ignored or the illumination change is not obvious.(2) Relying on the principle of region covariance descriptor, but with a probabilistic framework, we introduce an elegant way to integrate covariance descriptor into Mont Carlo tracking technique for object tracking. The advantages of particle filter and multiple features of region covariance descriptor entitle us better competence to handle object tracking within complex environment, as well as partial and completed occlusions of the tracked entity over a few frames. The experimental results show that region covariance based particle tracker outperforms CAMSHIFT tracker and color based particle tracker within complex environment because the region covariance descriptor has the ability to discriminate the object from others. And our tracker also better handle occlusions when comparing with region covariance descriptor based local search tracker. Because the proposed algorithm has adopted prediction-correction scheme to estimate the dynamic state and search the object in the predicted candidate region instead of local or global search, it is effective to find interest entity when it comes out after occluded for a few frames.(3) By applying the proposed method into robot base, we complete a system of mobile robot which can actively track the object initialized. The robot adopts face detection algorithm to initialize the object and then controls the robot base and the orientation scheme of the camera based on the object tracking method. The robot can successfully track the person with a uniform speed of 3cm/s. This system offers a stability foundation for household intelligent robot for surveillance and service.