Recognition of Human Body Posture from a Cloud of 3D Data Points using Wavelet Transform Coefficients

This paper addresses the problem of recognizing a human body posture from a cloud of 3D points acquired by a Human body scanner. Motivated by finding a representation that embodies a high discrimination power between posture classes, a new type of features is suggested, namely, the wavelet transform coefficients (WTC) of the 3D data points distribution projected on the space of the spherical harmonics. A Feature selection technique is developed to find the features with high discriminatory power. Integrated within a Bayesian classification framework and compared with other standard features, the WTC showed great capabilities in discriminating between close postures. The qualities of the WTC features were also reflected on the experiment results carried out with artificially generated postures, where the WTC got the best classification rate. To the best of our knowledge, this work appears to be the first to treat the posture recognition in the three-dimensional case and to suggest WTC as features for 3D shape. 1. Recognizing a 3D Human Body posture: Why ? The recent years have seen the emergence of human body scanners capable of capturing the whole shape as well as the appearance of the human body (HB). New perspectives were opened for the exploitation of this technology in various sectors. In entertainment , scans of real persons can be mapped to generic models and then integrated in video games, TV or cinema production [1]. In clothing industry, human body scans can substitute the real person for extracting measurements [2, 3, 4]. Data bases of HB scans can be useful for medical and anthropological surveys [5]. Many of these applications need decomposing the body shape into surfaces corresponding to the different parts of the human body, namely the head, the upper arms, the lower arms, the upper legs, the bottom legs and the torso. There have been some attempts to segment automatically human body shape [2, 3, 4] however these works treated the particular case of a standard posture (Figure 1.(a)). Their techniques were restrictive and cannot handle arbitrary postures. Our believe is that tackling the segmentation of hu-