Human pose estimation is of great importance in computer vision,which refers to the detection of human body keypoints.Accurate human pose estimation is a key step toward understanding people behavior in images and videos.It serves as the basis of action recognition,gait recognition,human-computer interaction and other higher-level tasks.Human pose estimation also has wide applications in many fields,including video surveillance,computer-assisted diagnosis,motion sensing game and virtual reality.In this paper,2D single-person pose estimation,multi-person pose estimation and video-based pose estimation are studied progressively based on deep learning.1.The interference of occlusion,illumination and complex background make challenges for single-person pose estimation.To address this issue,a pose estimation network with high resolution feature maps and large receptive field is proposed,which not only detects accurate keypoint locations,but also models highorder spatial relationships.Moreover,this paper presents a general method for online hard keypoints mining.Experiments and visualization analysis demonstrate the effectiveness of the proposed network.2.To capture the association between keypoints in bottom-up multi-person pose estimation,this paper introduces limb detection heatmaps as a representation of body keypoints association,which are simultaneously learnt with keypoint detections.Focal loss is adopted to address the data imbalance in human pose estimation.In addition,an inference optimization method is proposed to reduce redundant computation and improve inference efficiency.3.Simply applying a pose estimation framework designed for single image level to videos could lead to missing keypoint detections and false pose estimations due to motion blur and occlusion introduced by video frames.To address this issue,this paper proposes to utilize the keypoints from the nearby frames using temporal information expressed in optical flow.Visualization analysis demonstrates the effectiveness of this method. |