Statistical HOG on Multi-temporal Depth Motion Maps Approach for Human Action Recognition

Abstract

This work proposes an efficient spatiotemporal compact descriptor for action representation from depth map sequences. The feature descriptor is intended to resolve the problems of distinguishing different posture shapes with temporal order. The proposed work is composed of three phases. In the first phase, a depth sequence is partitioned into three non-overlapping temporal depth parts, which are utilized to produce three depth motion maps (DMMs) to capture the shape and motion cues leading to a multi-temporal DMMs representation. In the following phase, the Histogram of Oriented Gradients (HOG) is adopted from DMMs. Time-frequency statically features then extracted from DMM-HOG descriptor and concatenated in order to feed L2-CRC in the last phase. Comprehensive experiments on the known datasets clarify how the proposed approach exceeds action recognition related approaches. Experimental results achieved 97.93% and 95.97% for MSR Action3D, MSR Gesture3D respectively.