Reasearch And Implementation Of Learing Discriminative Action Pattern For Activity Recogintion

Human activity recognition is a significant task in computer vision and widely proved practical in many fields,such as video surveillance,video analysis and man-machine interaction.The advent of depth sensors greatly promotes the study of activity recognition.Being insensitive to illumination and occlusion,depth sensor can capture depth sequences containing abundant sufficient structural information and motion cues.This research is conducted based on the depth sequences captured by such depth sensors.This paper introduces a new framework where low-level semantic patterns are adopted to effectively study discriminative action patterns for efficient activity recognition.Thus,the research aims to solve these three problems:(1)The representation of action patterns.(2)The learning of discriminative action patterns.(3)The usage of discriminative action patterns in activity recognition.In this paper,each depth sequence is adaptively divided into several spatial-temporal action cube pool of each activity.The description of action cubes is defined as the action pattern.The vector aligning and feature pooling are used to integrate surface normal vectors of point clouds to represent action patterns.The spatial constrained clustering is used to cluster the sets of action patterns into several disjoint clusters.The Max-Margin Multiple-Instance Learning is used to mine the discriminative action patterns.The training set is iteratively utilized to learn a multi-class classifier which is used to select discriminative action patterns.In this paper,the descriptors of the depth sequences are constructed from the response distribution of the discriminative action patterns.Finally,two benchmark datasets,MSRAction3D dataset and MSRDailyActivity3D dataset,are used to show the validity and good performance of the framework.