Single-stage Human Pose Estimation Based On Deep Learning

Human pose estimation is one of the core tasks of computer vision,and has important scientific research value and practical application value.Human pose estimation plays an important role in promoting downstream tasks: gait recognition,pedestrian tracking,and pedestrian pose tracking.At the same time,there is a wide demand for human body pose estimation in the fields of intelligent monitoring,virtual reality,and athlete-assisted training.In recent years,with the application of deep learning technology in the field of computer vision,human body pose estimation has made some milestone progress.In general,the existing algorithms can be divided into two categories.The first category is top-down algorithms.These algorithms need to first use the object detection algorithm to convert the multi-person pose esttimation problem into a single-person pose estimation problem,and then the results obtained map back to the original image;the second type is the bottom-up algorithm,which first detects all human key points in the image,and then groups the human key points to get the human pose.Compared with the existing two types of methods,this paper proposes a single-stage human body pose estimation algorithm based on object detection,which introduces a structured representation of human body pose,and also designs a human body center sampling method to assign positive and negative samples.In order to achieve single-stage human body pose estimation,the algorithm in this paper performs dense prediction on feature maps.In order to improve the accuracy and solve the misalignment of pose and features,a pose proposal module and an adaptive pose feature alignment module are designed.Finally,a large number of experiments were conducted on a large public benchmark data set,which verified the effectiveness of our proposed algorithm,and achieved good results both in performance indicators and visual effects.