| 英文摘要 | With the growing requirement for safety, more and more cameras are utilized for visual surveillance. To keep computer aware what happens in the visual field automatically, motion and action understanding has long been the research focus, ranging from computer vision, pattern recognition to cognitive science. In this thesis, we concentrate on activity recognition and analysis in dynamic scenes, with focuses on motion feature extraction and representation, and generative modeling for object activity. Specifically, the following topics are addressed in the thesis: 1.Shape-based activity recognition using R transform: In activity recognition, object silhouette is commonly used to describe the spatial information which provides sufficient posture variance but with high dimensionality. Thus one key issue in activity modeling is to find an efficient feature descriptor with lower dimensionality, while keeping the robustness of the system. Consequently, R transform is adopted as a novel feature descriptor to represent the posture in each frame of an action sequence. Compared with other feature descriptors, R transform has low computational complexity, and is robust to frame loss, disjoint silhouette and holes in shape. Promising performance is achieved with the proposed method. 2.Interactive action analysis based on PCA-HMMs: In the case of multi-people activity recognition, much information is needed to represent low-level features. However, the high dimensionality of feature vectors for HMMs usually leads to covariance matrix singularity. In our solution, an improved inference scheme of HMM is proposed. Our approach distinguishes itself from the standard HMM work in that it uses the parameters with reduced dimensionality by PCA for recognition. Compared with traditional HMM, PCA-HMM gets higher recognition speed, and the performance gets comparatively better with the increase of the number of objects. 3.Group activity analysis based on statistical shape theory: Group activity analysis is challenging because of the large number of objects and occlusion among them. To tackle this problem, a system identification approach is proposed and applied in traffic surveillance. In this scheme, landmark points, instead of the complete trajectory are used to describe the dynamic information. Based on the landmark points, curves are formed and statistical shape theory is used to extract group activity features from the curves. Finally ARMA (Autoregressive and Moving Average) is adopted for feature learning and activity identification. As a robust scheme for activity analysis, the proposed solution is invariant with camera zoom and pan. |
修改评论