| Human detection and action recognition are hot topics in the field of ComputerVision and Pattern Recognition. The purpose is to establish such a system, which candetect and understand human appearance, movement and behavior, that is, to providea new human-computer interaction.In this paper, we did research both on human detection and recognition of humanbehavior. At present, immense amounts of research has been done on single-viewhuman detection. But the main problem is to solve the situation of occlusion body,while the method of human detection of multi-view is not common. In this paper, wepropose a multi-view body structure-constraints human detection method，and caneffectively solve the problem of multi-view human detection. As for the recognitionof human behavior, we presents information-based on the human skeleton Kinectcamera to obtain the behavior recognition method based on dynamic time warping.This method has a better adaptability to the change of views.The main contents include:(1) Human detection: In order to better carry out information fusion betweenhuman perspectives. We uses human detection model which based on blockinformation to obtained body information from different views.(2) Information fusion among different views: For the information fusion ofdifferent views, we leverage spatial affine transform to correlate the overlappingregions in different views. To overcome the challenging of object corresponding inmulti-view environment caused by partial occlusion and multiple target existence inneighborhood, we formulate the model of maximum a posterior (MAP) for multipleobject correspondence by taking advantages of the body structure constraints.(3) Human action recognition: With the skeleton information extracted fromKinect camera, we take use of dynamic time warping (DTW) algorithm and usetemplate matching to get right classification.The innovation of this article is to propose a multi-view bodystructure-constraints human detection method and can effectively take use ofmulti-views’ information. The experimental results show that the propose method can achieve good detection result. For human action recognition, experiments wereconducted on the Microsoft Action3D, Microsoft Daily Action3D data sets fortesting and TJU respectively, which indicate that the propose method is effectively. |
