The Study Of Human Pose Estimation In Monocular Video

The target of human pose estimation in monocular video is to accurately estimate thebody part positions in3D, called pose, of a moving human body in monocular video. Due tothe inherent difficulties and wide applications, this topic has been receiving increasingattention from computer vision researchers.This thesis mostly focuses on the problems of moving object detection and human poseestimation in monocular video, and the main contributions are as follows:Firstly, to detect moving object in static background without prior knowledge of thebackground, this thesis employs a method for moving object detection based on Three FrameDifferencing and Codebook Model which generates the human body silhouette as the input ofpose estimation. The algorithm can detect and extract the human body silhouette.Secondly, aiming at detecting object when the camera and object motions are mixedtogether, this thesis proposes an approach for moving object detection using continuoustracking optical flow in image sequences. First, it tracks feature points between first andsecond image frames using the pyramidal Lucas-Kanade method and then groups thesetracked points based on their motions. Next, it continues to track the grouped feature pointson subsequent image frames. For these points in each group, the maximum aggregate will bereserved and undersize groups must be removed in the continuous tracking process, until theonly one group is chosen. All the feature points belong to background in this group and it canbe applied to compensate for background motion. Finally, it detects the moving objectemploying frame difference method. For subsequent image frames, the feature points in thechosen group are tracking continuously. The experiment results demonstrate that the methodimproves both the accuracy and the speed of moving object detection.Finally, for the model-free pose estimation techniques, this thesis adopts one method toestimate human pose based on matching the human body silhouette. It uses the ShapeContexts descriptor to describe the silhouette shape, and figures out the best matching sample from the sample set and makes its corresponding3D pose parameters as the result ofestimation. This can make the pose estimation in monocular video come true.