Indoor Human Action Recognition Based on Computational Model of Visual Perception on Neocortex and Depth Information
Abstract：Human action recognition in daily living has been actively studied due to its relevance to a large variety of applications, such as video surveillance, video retrieval, human-computer interaction. The difficulties of action recognition come from several aspects including its complexity and variations of human action, uncertain settings. On the topic of representation learning for human action, we have conducted extensive studies on the nature of Intelligence with the neocortical structures and connectivity, and strategies of sensory coding. It is first time to define this mechanism as hierarchical models based on sparse coding and redundant filtering, and propose a novel computational model for visual perception on neocortex, and then introduced into the human action recognition with the depth (RGB-D) sensors. 1) Mid-level representations for action recognition are presented based on mathematical algorithm of slow feature analysis (SFA) and bionic algorithm of sparse distributed representation (SDR), respectively. 2) Multi-class SVM classifier or long short-term memory model is trained for individual action classification. And-or graph promotes the action recognition in the scene with individual-scene interaction. Finally, a prototype system of video surveillance is established for indoor human action recognition in daily living.
Keywords: Action recognition; Visual perception; Depth information
Contact:
JIA Lihao
E-mail: lihao@ia.ac.cn