Research And Application Of Real Time Multi-Person Pose Estimation For Surveillance Video

Computer vision-based human pose estimation is to obtain the position of human joints and the interconnection between joints from a monocular RGB image.Human pose estimation has a broad use in areas such as intelligent surveillance,human-computer interaction,virtual reality,motion analysis and so on.With the development of deep learning algorithms,human pose estimation based on deep learning have gradually replaced traditional algorithms based on Picture models,both the accuracy and speed of pose estimation have been improved.However,the existing human pose estimation based on deep learning still faces great challenges in real-time processing.The main goal of this paper is to realize real-time multi-person pose estimation without reducing the accuracy,and promote the application of human pose estimation in intelligent surveillance.The method of human pose estimation in this paper is based on deep learning,and the bottom-up method is used to estimate multi-person pose.The neural network firstly predicts the position of all joints in the image,then connect joints to form a graph and the pose is estimated base on the graph optimization method.The runtime of pose estimation algorithm,which is based on bottom-up method,is independent of the number of people.The neural network in this paper contains two branches and six stages.The multi-stage can expand the overall receptive field and improve the accuracy of joint prediction;For the two branches,one branch predicts the position of joint.The other predicts the spatial relationship between joints.In order to represent the spayial relationship between joints,a limb affinity zone is proposed in this paper,which consists of a series of unit vectors that can simultaneously describe the position and direction of the limb.The limb affinity zone can effectively sparse the redundant connections between the joints so that reduce the complexity of the graph optimization.In addition,the human pose estimation algorithm is applied to the detection of abnormal human behavior in the surveillance scene in this paper.To analysis the abnormal human behavior,KCF algorithm is used to track the joints,and then calculate the motion speed of each joint.Determining whether there is abnormal behavior by comparing the relative motion speeds between different joints on the same person.The real surveillance video is used to verify the algorithm,the training samples are video frames with different time periods,different resolutions,different crowd density and different weather conditions.The training results show that the average prediction accuracy of each joint in the surveillance scenes reaches 0.84,which is better than current existing deep learning methods.Three different resolution videos,1920×1080,1280x720 and 720×480,are used to analyze the efficiency of the algorithm.For the case that neural network take every frame of the video to estimate the human pose,it takes 23 frames/s,30 frames/s and 46 frames/s,respectively.While the algorithm takes 35 frames/s,41 frames/s,and 60 frames/s,respectively if KCF tracker is used to track joints in the internal frame of video.Thus show that the neural network proposed in this paper satisfies real-time human pose estimation,and can be used in processing real-time surveillance stream.