Human behavior analysis attempts to detect, track and identify people, and more generally to interpret human behaviors, from image sequences involving humans. Human motion analysis is currently one of the most active research topics in computer vision. This strong interest is driven by a wide spectrum of promising applications in many areas such as smart surveillance, perceptual interface, virtual reality, content-based video retrieval, and etc. The thesis is mainly focused on the following aspects: (1) Local spatio-temporal feature representation; (2) Human behavior sequence segmentation; (3) Unsupervised behavior clustering. The main contributions are: 1. A review of the state of the art in behavior analysis is carried out, and the methods on behavior analysis are broadly divided into 4 classes: (1) Action representation methods; (2) Behavior segmentation methods; (3) Unsupervised/semi-supervised clustering methods; (4) Behavior modeling and recognition methods. 2. The entropy-based image salient point detection algorithm is used in human action analysis: (1) A novel approach, which uses the action repetition and consistency as the two cues of video contents, is proposed to cluster human actions automatically from video sequences. (2) A novel local spatio-temporal feature representation is introduced. 3. A novel unsupervised algorithm for behavior sequence segmentation is proposed. The algorithm consists of the following steps: (1) The video sequence is coarsely segmented into equal length subsequences with overlapping time window. (2) Segmental-DTW is used to find out matching behavior clips between pairs of video subsequences. (3) The similarity between behavior clips is represented by an adjacency graph, and an efficient graph clustering algorithm is used to generate behavior clusters. The algorithm, based on a coarse-to-fine strategy, is able to satisfactorily segment behavior sequences and cluster typical behavior patterns. The segmentation results could be used for further behavior modeling and recognition. 4. A novel human action sequence segmentation and recognition algorithm is proposed. Taking into account of the continuity of human actions, we argue that the problem of human action sequence segmentation should not be isolated from that of the action representation and recognition. In our work, human action sequences are clustered by constructing a new bipartite graph from the co-occurrence matrix of action sequences-prototype features, and a theoretical analysis is carried out on its applicabilities. Based on MAP-MRF as well as the prior knowledge of action clusters and the distances between prototype features in low-dimensional spaces, an automatic human action sequence segmentation and recognition algorithm is obtained.
修改评论