英文摘要 | Visual analysis of human motion is currently one of the most active research topics in computer vision. It aims to recover body poses and motion parameters from static images or video sequences. The recovered data, used for pose recognition, semantic analysis and behavior understanding, have a wide spectrum of promising applications in many areas such as virtual reality, smart surveillance, advanced perceptual interfaces, motion analysis, and model-based coding. In recent years, although visual analysis of human motion received increasing attention from both academia and industry, many theoretical and technical problems remain open. This thesis focuses on an important subject in this field, i.e., model-based tracking of walking people, which not only involves many issues of low-level vision but also provides motion data for high-level visual analysis. Model-based tracking of walking people is a general framework for people tracking. Under such a framework, we analyze the important modules (including learning and representation of prior knowledge, pose evaluation function, initialization, search strategy, and so on), describe some novel algorithms, and draw some useful conclusions. Our contributions are summarized as follows. (1) A compact motion model is learnt from a volume of training examples. The model, represented as Gaussian distributions, plays an important role in prediction and initialization. We also carefully analyze the human motion constraints: intervals of joint angles and dependencies of neighboring joints. The former are derived from confidential intervals of joint angle distributions that are modeled as mixture Gaussians. The latter are represented by conditional distributions whose parameters are learnt from training data. (2) We propose a robust approach to motion detection that is applied to extraction of features of edges and region information. Then both features are combined into the pose evaluation function to obtain accuracy and robustness. (3) People tracking is an optimization problem of high dimensionality. We decompose it into two sub-problems: estimation of global position and refinement of joint angles. As to the latter, we propose an effective approach to recursively refine each joint separately. This approach is based on the spring model and rotation kinematical equation. (4) To avoid the deficiencies of the above approach, we also track people in a probabilistic framework using a particle filtering. According to particle filtering, we emphasize on the initialization and dynamic model. We use the spatio-temporal information of the first N frames and prior knowledge of human motion to initialize the body pose. Then tracking history, motion model and motion constraints are fused to design our dynamic model. |
修改评论