Microsoft Kinect is a low-cost range-sensing device that captures depth maps; its Windows software development kit (SDK) enables programmers to extract human postures (the various positions that the human body can assume) from such depth maps in terms of 3D skeletal models with 20 body joints. In this paper, the authors propose an improvement to the posture description and skeleton formation algorithm in the Windows SDK.

In the proposed algorithm, the process of skeleton formation grossly mimics the behavior of the Windows SDK. It starts with a calibration stage that converts the depth map to a 3D point cloud. The authors then engage a learning algorithm to label 31 different body parts, describe the skeletal joints, and compose a posture. Translation-invariant depth features of the body parts are used in a random decision forest to recognize “belongingness,” that is, the association of every pixel with a likely body part. This per-pixel information is finally collated to propose a skeletal joint for 3D joint positions as the output. The collation algorithm is based on local modes of probable 3D positions, and hence is much less affected by outlying pixels, which can severely adversely impact the global 3D center-of-mass approach used in the Windows SDK.

The authors demonstrate the quality of the skeleton model and the extracted human posture using a 3D light emitting diode (LED) cube, which helps with physical visualization.

This paper deals with several steps involved in skeleton extraction from depth maps. However, its independent contribution is limited to local optimization in the identification of likely body joints. As a result, the paper assumes familiarity with the skeleton extraction algorithms. This restricts its potential readership to groups of researchers focused on Kinect and related low-level information synthesis strategies.