Research On Robustness Of Human Keypoints Detection Algorithms

Human keypoints are essential for describing and analyzing human behavior,and human keypoints detection is also the basis of many computer vision tasks.Generally speaking,human keypoints detection needs to detect keypoints of key body parts,such as face and skeleton,which respectively correspond to two basic sub-tasks of facial landmark detection and human pose estimation.Facial landmark detection aims to localize representative facial semantic feature points,such as eyes,noses,mouth and cheek,and is a prerequisite of many automatic facial analysis tasks,such as face recognition and facial age estimation.Human pose estimation aims to locate representative body parts,such as arms,torsos,hips,ankles and knees,and plays a fundamental role in many related fields,such as activity recognition and gait recognition.Therefore,the research on human keypoints detection has great application value,and this paper studies two aspects of human keypoints detection from the perspective of robustness of algorithms:For the task of facial landmark detection,the annotated landmarks in the dataset usually deviate from the ground truth landmarks.However,existing facial landmark detection algorithms regard the manually annotated landmarks as precise hard labels,and do not take the effects of inaccurate annotated landmarks in the training set into consideration,lacking the robustness against the annotated noise.This paper proposes the facial landmark detection algorithm by label distribution learning.The algorithm proposes to associate a Bivariate Label Distribution(BLD)to each landmark,and covers the neighboring pixels around the original manually annotated point,alleviating the problem of inaccurate landmarks.Experimental results show that the proposed method performs better than the compared state-of-the-art algorithms.Furthermore,the proposed method appears to be much more robust against the landmark noise than other compared baselines.For the task of human pose estimation,although existing deep learning based approaches have achieved significant progress by fusing the multi-scale feature maps,they have not fully enhanced the information of feature maps that is more effective for the pose estimation.Therefore,there still exist problems against the robustness of keypoints detection in complex scenes,such as invisibility,occlusion and large poses.This paper proposes the multi-person pose estimation with enhanced channel-wise and spatial information.The algorithm firstly proposes the Channel Shuffle Module(CSM)to promote cross-channel information communication among the pyramid feature maps,and then adopts the Spatial,Channel-wise Attention Residual Bottleneck(SCARB)to highlight the information of feature maps both in the spatial and channel-wise context.Experimental results show that the proposed method achieves the state-of-the-art results,and appears to be much more robust against the keypoints detection in complex scenes.This paper consists of five chapters.The first chapter introduces the research status of human keypoints detection and the problems to be solved.The second chapter introduces the specific definitions and related algorithms of two human keypoints detection subtasks(i.e.,facial landmark detection and human pose estimation).The third chapter introduces the facial landmark detection algorithm by label distribution learning,and the detailed experimental results,robustness analysis.The fourth chapter introduces the multi-person pose estimation with enhanced channel-wise and spatial information,and the experimental results,robustness analysis.Finally,the last chapter are the summary and further prospects.