Research On Human Pose Estimation In Monocular Image And Video

Human pose estimation is a hotspot and difficult issue in the field of computer vision.It has wide range of application prospects in many fields,such as intelligent monitoring,advanced human-computer interaction,image and video retrieval,virtual reality and motion analysis.There are several key problems to be solved in the pose estimation in monocular images and videos.For example,the lower part searching efficiency due to large searching space,the false detection leaded by the diversity of human pose,the loss or false detection caused by occlusion,the higher inference complexity of video pose estimation structure and so on.In order to address the above four problem,many approaches are proposed including human parts searching algorithm with MRF superpixels part label,diversified pose representation on multi-model structure,occlusion level for occluded part model,pose inference with composite model in video.The main contents of the thesis are as follows:The body parts searching algorithm based on MRF superpixels label is proposed to improve the efficiency of part scanning process.The MRF superpixels labeling approach and square parts detector is introduced to scan the superpixels for parts match.Firstly,superpixels is used to segment image into several blocks.And then the MRF part model on superpixel is employed to label superpixels for the body part;secondly,the part search is guided by the superpixels,and the superpixel center is identified as the matching center of the square part detector.Finally,the deformable part model is used for human pose estimation.The experimental results on dataset IP and LSP show that human body parts are successfully labeled in the approach and decrease the searching time and estimate pose efficiently.A human body model suitable for human body structure is studied aiming at the problem that the pose estimation is affected by the diversity of human poses.A human pose estimation method based on GCT(Global-Constellated-Tree)model is represented.Firstly,the initial candidate for global human orientation is obtained by multi-global detector for human body.Secondly,constellated multi-model is proposed to depict the relationship between the human global orientation and diversified body parts.Thirdly,the global-constellated model and tree model is fused together in order to better describe the distributed relationship between adjacent parts.The experimental results on dataset IP and LSP show that the method improves the accuracy of pose estimation,especially for diverse pose.An occluded part modeling method with occlusion level is proposed in order to comprehensively consider the parts detection and human body structure.The model with occlusion level denotes the parts detector and relationship between adjacent parts in a variety of occlusion.Firstly,the occlusion level is defined as the occluded degree of human body parts,which is acquired by calculating the ratio of occlusion and orientation of parts;then,a more robust human body part model with occlusion level is established in order to decrease the interference caused by occlusion,and the deformable part model combined with occlusion level is introduced to indicate the relation between neighboring parts;finally,the proper human pose is estimated according to both the part model and deformable model.The experimental results on IP and LSP datasets show that our method improves the overall accuracy of the pose estimation,especially in the case of occlusion.A human pose inference method based on CES(Composite-Elementary-Symmetric)model is proposed in view of the low accuracy of pose estimation in video and the high complexity of inference.Firstly,a composite part model which represents rigid body parts is proposed aiming at simplifying the complexity of multi-frame inference.Secondly,the elementary part model is proposed to be linearly connected as composite part in order to improve the detection rate of little part of the human body in single frame.Finally,the symmetric parts model which uniformly represent several symmetric body parts is introduced into the composite model to improve the efficiency of inference.The model is used to reduce the double-counting of human body parts and optimize the inference and improve the overall efficiency of pose estimation.The results on the standard dataset(Outdoor,Human Eva-I and N-Best)show that the approach improves the accuracy of pose estimation and enhances the efficiency of inference in video.