| With the development of deep learning,human pose estimation and tracking technology plays a significant role in human body modeling,behavior recognition,and safety control.Especially in the field of surveillance,we can utilize pose estimation and target tracking systems deployed on multiple platforms to liberate labor,greatly improving efficiency.However,the current state-of-the-art human pose estimation performs well under favorable conditions such as unobstructed and clear images,but it cannot handle situations with multiple people in videos that have motion blur,video defocus,or pose obstruction.It is because it cannot use the spatiotemporal information between consecutive video frames.In this thesis,a new multi-frame human pose estimation framework is proposed,which utilizes the rich temporal clues between video frames to facilitate key point detection.The specific work includes the following points:1.This thesis is based on the HRNet feature extraction network algorithm.The features extracted by the HRNet backbone network are fed into a pose merge network,and the prediction of the current frame is corrected based on the information obtained from the network fusion.The pose merge network in this thesis is divided into three parts.The first part of the network performs pose merge on multiple spatial images to obtain spatial-temporal information about joint movement.The second part of the network extracts features by grouping them according to joint correlation to reduce interference from unrelated joint information.The third part of the network uses a multi-scale perception method to extract motion features to correct the prediction results.The thesis proposes a high-performance network and a light network.The high-performance network performs well compared to mainstream single-person pose estimation networks on the Pose Track2017 and Pose Track2018 datasets.The light network achieves a nearly four-fold improvement in inference speed compared to the high-performance network with multiple batches2.This thesis designs and implements the human body pose assessment and tracking system,and completes the deployment of the pose merge network and the two-stage target tracking network.The Tensor RT framework is used to optimize the model algorithm and solve the problem of algorithm model landing.The system achieves a balance between reducing the model size and increasing the model computation speed while controlling the accuracy loss within an acceptable range.It also supports multiple AI capabilities and realizes efficient inference with composite functions.This thesis introduces spatio-temporal information into the research of video-based human pose estimation algorithms,optimizes the algorithm structure,and achieves improved accuracy.The implemented system performs well in the use of algorithms,and it also has certain practical value. |
