Study On Human Action Recognition Algorithm

Human action recognition is a new research topic in machine vision and artificial intelligence, and the objective is to detect and recognize human actions from videos so that the computer system is able to understand human behaviors and make further semantic description of the scene. Analysis of action based on video, mainly consists of two parts:human motion region detection and action recognition. Human motion region detection is to detect human motion region in video by using the method of moving target detection. Action recognition is to extract and describe the features in the human motion region, and adopt the suitable classification method for training and classification. There are challenges in both of these two research aspects. In the aspect of human motion region detection, the main challenge is to detect human motion region with different moving speeds in complex background clusters, and under different illumination changes. For human action recognition, action is composed of3D information, cannot be well described with the conventional image processing methods, so the feature extraction and description and classifier design is full of challenge.The main work of this thesis includes the following aspects:1. For the redundant information of the video, this thesis proposes a method to remove redundant frames of a video. The removal of video redundancy can reduce the amount of calculation, and avoid the interference of redundant information. In order to accurately extract human motion region, this thesis adopts the method of background modeling and frame difference method are combined, constructs a human motion region detection method2. For the feature extraction and description, this thesis develops a3DHOG feature based on human moving region. In addition, adopts two kinds of good performance features:global descriptors based on Gabor filtering on3D frequency domain; local descriptor based on spatio-temporal interest point. These three features are described with the bag of words model, are applied to action recognition system.3. In the part of multiclass classification, this thesis deeply studies the theory of support vector machine algorithm, and the multiclass classification strategy. For the three features, a multiclass classifier of multi feature fusion is designed, which makes up the defect of single feature for fully describe the action, and the system recognition rate is improved. This thesis carry out the experiment in benchmarks database, experimental results demonstrate that fusion of multiple feature substantially outperforms a single feature, and the fusion of three kinds of features significantly improves the precision of recognition.