Implementation of levels-of-detail in Bayesian tracking framework using single RGB-D sensor

This paper propose a study of real-time human gait tracking system based on a cascaded particle filter implementation using Microsoft Kinect sensor. Our tracking system is combination of two different levels which processing both color and depth information. In the first level, we utilize color histogram to implement a coarse 2D region tracking. For the second level, we implemented two different depth feature extractions, i.e., spin image and geodesic distance for tracking extremities of the body. These two levels are combined to represent the basis for the full 3D human body tracking by obtaining a probable human body bounding box in 2D while the 3D position is obtained by incorporating available depth information. The relationship between motion and particle generation is modeled for tracking a person in the first level. For this level, we adopt the expected motion constraints for enhancing the distribution of particle at the importance sampling stage. In the implementation of the second level, 3D bounding box coordinate system is generated by synchronizing with RGB and depth video stream. In a coarse-to-fine cascaded concept, we used spin-image and geodesic depth information to track the extremities in the second level. State definition of the points of extremities is defined, and the depth-based particle filter is implemented for extremity point tracking.