Human action recognition based on key postures

Publication Type:
Thesis
Issue Date:
2009
Full metadata record
Files in This Item:
Filename Description Size
Thumbnail01Front.pdf4.68 MB
Adobe PDF
02Whole.pdf69.45 MB
Adobe PDF
Human motion analysis has gained considerable interests in the computer vision area due to the large number of potential applications and its inherent complexity. Currently, human motion analysis is at an early stage. Its final aim is to generate an easy understanding, high level semantic description in a given scene. Human action recognition is an important step to the final aim of human motion analysis. Human Detection Human detection is part of the field of human motion analysis. The thesis looks at human detection. The thesis proposes a method using histogram of angles to discriminate pedestrians from vehicles. This proposed method is encouraged by the reality that humans are non-rigid objects, An angle formed by the centroid point and two bottom points for a human changes periodically while the angle for the vehicle is relatively static. In this part, this thesis also presents an approach to detect humans in static images. The thesis proposes an approach which uses human geometric features to fulfill the task. Human Action Recognition The thesis focuses on human action recognition. The thesis proposes what will be called a key postures based human action recognition approach. As we have known, human actions can be well described by a few important postures (called key postures) which are significantly different from each other and all other postures can be clustered to these key postures. Therefore, these key postures can be used to represent and to infer the corresponding human action. The benefit of using key postures to represent human action is to reduce computational complexity. The thesis proposes two methods for human action recognition based on key postures. One is a human action recognition based on shape features and the other one is action recognition based on Radon transforms. Both methods follow three steps to achieve action recognition. These steps are video processing, key posture extraction and action recognition. A two-step approach is proposed to extract key postures from preprocessed action video. These two steps are coarse selection and fine selection. Feature extraction and representation are discussed in both steps. After key postures are extracted from a video, key posture sequences are used to represent human actions. Each key posture sequence is regarded as an action template. In order to compare two action sequences, Dynamic Time Warping (DTW) is applied to determine the distance between the two action sequences. In the second method, in order to obtain key postures, the action sequences are extracted from the preprocessed silhouettes using Radon transforms. Then, an unsupervised cluster analysis is applied to Radon transforms to identify the key postures for each sequence. Such key postures are used in the subsequent training and testing procedure. Several benchmark classifiers are used in this work for action learning and classification. Author's Publications This thesis covers the research results conducted by the author while undertaking for the degree. Most of the results have been published in research papers in refereed publications which are listed in Author's Publication for Doctor of Philosophy (PhD).
Please use this identifier to cite or link to this item: