Action recognition from point cloud patches using discrete orthogonal moments

Abstract

3D sensors such as standoff Light Detection and Ranging (LIDAR) generate partial 3D point clouds that resemble patches of irregularly-shaped, coarse groups of points. 3D modeling of this type of data for human action recognition has been rarely studied. Although 2D---based depth image analysis is an option, its effectiveness on this type of low-resolution data hasn't been well answered. This paper investigates a new multi-scale 3D shape descriptor, based on the discrete orthogonal Tchebichef Moments, for the characterization of 3D action pose shapes made of low-resolution point cloud patches. Our shape descriptor consists of low-order 3D Tchebichef moments computed with respect to a new point cloud voxelization scheme that normalizes translation, scale, and resolution. The action recognition is built on the Naive Bayes classifier using temporal statistics of a `bag of pose shapes'. For performance evaluation, a synthetic LIDAR pose shape baseline was developed with 62 human subjects performing three actions - digging, jogging, and throwing. Our action classification experiments demonstrated that the 3D Tchebichef moment representation of point clouds achieves excellent action and viewing direction predictions with superb consistency across a large range of scale and viewing angle variations.