Tri-tracking Algorithm Research For Video Tracking Based On Kinect

As the key technology of the fields of computer vision, visual tracking plays an important role in intelligent control, human-computer interactions, medical diagnosis and other fields. However, robust tracking is still a challenging problem, due to the appearance variability of the object caused by rotation posture and external factors such as illumination, occlusion and local blur.3D camera Kinect is used in our system. The color and depth images can be obtained by Kinect simultaneously. Based on the depth images from the Kinect, we can segmentation the area nearby the target from the other background, so to reduce the interference of the background. Based on the segmentation images, we present a novel tri-tracking framework mixing Incremental PCA model, Online Adaboost model and Hausdorff distance model for robust tracking. This tracking framework employs discriminative and generative model based on online semi-supervised learning. In the tracking process, we only need to initialize the first frame manually and the other frames are proceed automatically, with updating the models to capture the changes of the target appearance and the background.The main functions of our tri-tracking framework are as follows:Firstly, in this paper, the tracking robustness will be increasing by the combination of depth images and color images. Secondly, a tracking framework that combining generative model and discriminative model is proposed, so we can get the model not only the object but also the background, the two kinds of models can make up the shortcomings of the other. Thirdly, by the introduction of the three models, we can deal with the situation when the two models have the opposite results in co-tracking framework. Finally, we propose a way to mixture of multiple models based on their discriminative ability.We test a lot of challenging videos, experimental results demonstrate that our algorithm is more robust in dealing with the motion blur, illumination, complex background, the scale changes and occlusion.