Research On Depth-based 3D Human Pose Estimation

In recent years,with the development of computer vision and depth camera,the 3D human pose estimation has made great progress,showing its value in many application fields.Currently,depth-based human pose estimation datasets are relatively small and lack of diversity.Classic pose estimation algorithms based on heatmap regression and 3D deep learning also have problems such as non-adaptive towards joints and low runtime efficiency.Aiming at these problems,this paper proposes a series of improvements in knowledge transfer and algorithm runtime efficiency.Existing public depth-based human pose estimation datasets are small in scale and insufficient in diversity,and the existing pose estimation algorithms cannot make good use of large-scale annotated RGB datasets,this paper proposes a method based on cross-modality knowledge transferring and multi-scale local refinement,by reducing the domain shift between the RGB domain depth domain,we can make better use of a large-scale RGB datasets,and prevent our model from over-fitting to the limited number of depth training samples.In order to select the local region in the " coarse-to-fine" human pose multi-stage pose estimation algorithm,we propose a multi-scale local refinement network to integrate richer scale-aware contextual information,which further improving the algorithm's performance.Heatmap-based human pose estimation methods are generally trained with non-adaptive ground-truth Gaussian heatmap for different joints and with relatively high computational burden.3D deep learning-based methods are difficult to train with costly voxelizing procedure,due to the large number of convolutional parameters,point-set based methods require some extra time-consuming preprocessing treatments(e.g.,point sampling).Towards these problems,this paper an anchor-based 3D articulated pose estimation approach for single depth image termed A2 J,Within A2 J anchor points are densely set up on depth image to capture the global-local spatial context information,and predict joint's position in ensemble way.The wide-range experiments demonstrate A2J's superiority both from the perspectives of effectiveness and efficiency.Aiming at the problems of low runtime efficiency of general model ensemble methods,this paper proposes a human pose estimation method based on deep negative correlation learning.By introducing the concept of negative correlation learning,we only need to modify a small parts of network structure,with the amended loss function,the proposed method achieves better performance while keeps the runtime efficient.