Research On Monocular 3D Human Pose Estimation

Monocular 3D human pose estimation has attracted significant attention in computer vision due to its widespread applications.However,it is a challenging task due to occlusion,viewpoint variance,and the ill-posed nature of back projection.In this thesis,we follow a standard two-step pipeline which first detects 2D joint locations and uses them to infer 3D pose.Furthermore,we combine the monocular 3D human pose estimation with content-based image retrieval,and propose a view-invariant human pose image retrieval framework.Our main contributions include:(1)Selection of deep learning network for 2D human pose estimation.In order to accurately obtain human joint position from monocular image,we compare two popular 2D pose estimation depth network models,CPM(Convolutional Pose Machines)and stacked Hourglass model,in both theoretical and experimental aspects.Experimental results show that stacked Hourglass model can better deal with the occlusion,so we choose it to detect 2D human pose from monocular image.(2)Since the exemplar-based method depends on the variety of the exemplar set,we propose a novel augmented exemplar-based algorithm.We use a strategy of matching and synthesis to implicitly augment the exemplar set for 3D human pose estimation,ensure that the augmented exemplar set has much more variety.The motivation of our algorithm is to well represent various poses in the real world with finite real exemplars.In our algorithm,we firstly matching the full-body and half-body poses,and synthesize virtual candidate poses using half-body exemplars.Then,we present an effective approach to select the best exemplar from candidate set to well match the detected 2D pose.We do a lot of experiments in Human3.6M dataset.Experimental results show that our method achieves competitive performance,especially for some complex poses such as 'sit down'.(3)Most existing human pose image retrieval is affected by viewpoint.In order to solve this problem,we propose a view-invariant human pose image retrieval framework.The framework use the result of the monocular 3D human pose estimation to extract view-invariant feature and measure the similarity of the human pose images.Experimental results on multi-view action database IXMAS show that our method can retrieve similar pose image that taken in different viewpoints,and improve the performance of human pose retrieval.