This work presents a robust method for realtime segmentation and tracking of moving objects using depth image sequences, which is insensitive to illumination changes. We propose a novel criterion in our quadtree split-and-merge framework and effectively solves the problem of segmenting objects in complex and cluttered scenes. We also introduce a plane estimation algorithm to cope with the indistinction of depth between the ground and objects standing on it. In order to better identify foreground objects and enhance the stability of the tracking, a depth-based background model is integrated into the system. Experiments in various scenarios demonstrate the high speed and excellent performance of our procedure. The output of the system will benefit further research on human pose estimation and activity recognition.
展开▼
