Vision Based Human Detection,3D Pose Estimation And Motion Analysis With A Single Camera

Computer vision serves for advanced Human-Computer interaction. Itsresearch content is that through the corresponding theories and algorithms,computers can reach and even exceed the standards of human visual percep-tion and cognition.Object detection, human pose estimation and face detection are threeessential issues in the field of computer vision. So far, object detection tech-nologies have been widely used in smart surveillance, security monitoringand detection. Human pose estimation attracts a lot of interests in virtual real-ity, somatosensory interaction, sports pose analysis and health monitoring.Face detection is the fundamental means for security entrance guard system,identity recognition and object tracking. In this thesis, we focus on these top-ics to study and discuss. It is not only attached with the foundation of thecomputer vision issues, such as feature extraction and classification andmodel description, but also faced to high-level semantics image understand-ing, such as human pose estimation and facial pose estimation. In recentyears, academia and industry have put in a lot of investigation on these hotresearch spots, and achieved remarkable results, which have a broad applica-tion prospects. (1)We proposed a pedestrian detection algorithm in public places basedon feature extraction and classification. For the difficulties in public places,such as occlusion and human posture diversities, we just select thehead-shoulder part of pedestrians as feature extraction target, and proposed aworkable solution based classification method SVM.(2)3D human pose estimation method under monocular camera. We pro-posed a novel method on monocular vision3D pose estimation. Monocularcamera causes a lack of depth information. So we attempted the annealedparticle filter method and made use of the continuity to reduce the depend-ency of3D motion on depth information. The initialization of motion isthrough manual calibration.(3)At the last section in our work, we proposed a novel and robust facedetection and pose estimation method. It contains two matching models, ap-pearance model and deformation model. The deformation model allows facialtopological changes due to different viewpoints. We implemented the methodbased on dynamic programming and search the optimal matching pattern inthe global scope. And we optimized the algorithm in efficiency with the ap-proach of parallel computing.