Action Recognition Via Global And Local Movement Patterns

With the rapid development of computer technology, the massive growth of image and video data has promoted more research and application works in computer vision, artificial intelligence, pattern recognition, machine learning and other related fields. As a high-profile research area in computer vision, human action recognition has been widely applied in intelligent video surveillance, patient monitoring systems, human-computer interaction, virtual reality, smart home system, intelligent security and athlete training, which has widespread value and prospect for application. However, classification of some human actions remains a challenging research problem due to background complexity, camera motion and variation in human appearance, posture and body size within class, etc.Traditional action recognition technologies based on interest points normally generate feature vectors by extracting low-level features (such as HOF, HOG and 3D-SIFT) to achieve the local information in video before constructing semantic bag-of-words model. The advantage of these technologies lies in no preprocessing of detecting moving objects. In that case, it will not be influenced by the results of image preprocessing, and the technologies are not sensitive to noise and interference factors as well. However it has disadvantage of high computational complexity, especially in the training process as it takes long time to build the dictionary model. In this dissertation, we propose a novel approach concerning more profile and structure information of the interest points, defined as movement patterns (MPs) in our approach, instead of extracting feature descriptors of interest points. The cardinal contribution of our work includes:1) With the interest points detected (Dollars detector), we propose a framework to discover the MPs of an action, including global movement patterns (GMPs) in the first layer and local movement patterns (LMPs) in the second layer, which can be used for a two-layer tree classification model 2) For GMPs, we detect the spatio-temporal interest regions and discover global patterns by predicting the spatio-temporal interest regions using Kalman filter, followed by dynamic time warping and classification. In the training phase, cross-word reference templates (CWRTs) is utilized to construct the reference GMPs.3) For LMPs, the inner interest point trajectory as well as weighted movement velocity in an interest region is extracted to produce the local patterns, followed by an adaptive clustering algorithm to create reference LMPs, which are used for classification in the second layer.In the training phase, a two-layer classification model is constructed with top layer including multiple unambiguous classifiers and ambiguous classifiers, and with bottom layer including multiple unambiguous classifiers built for the each ambiguous classifier in the top layer. Definition of unambiguous classifier:classifier including only one category of action. Definition of ambiguous classifier:classifier including multiple categories of similar actions. For a test action, GMP descriptor and DTW technology are used for recognition. If it is classified into the unambiguous classes in the top layer, the classification process is ended. If it is classified into the ambiguous in the top layer, LMP descriptor and template matching method are utilized for further classification in the bottom layer.The method in this dissertation is verified in the standard Weizmann and UCF dataset as well as in the multiple cameras fall dataset, achieving accuracy of 93.0%, 88.4% and 93.5% respectively. Experiments indicate that guaranteeing a high accuracy of recognition, the method in this paper outperforms the conventional bag-of-words approach based on multiple low-level features (such as HOF, HOG and 3D-SIFT) in computational complexity as well as in real-time capability. Finally, the influence of three different kinds of parameters on accuracy are analyzed.