Research On Human Pose Estimation Method And Development Of Motion Simulation Assisted Fitness System

Human pose estimation is an important part of many computer vision tasks,and plays an important role in the fields of human-computer interaction,video surveillance,and video pedestrian action recognition and analysis.Compared with static image-based human pose estimation methods,video-based human pose estimation has begun to receive more and more attention,because the correlation of time series can be used to produce better pose estimation performance.Estimation methods still face challenging problems such as human occlusion,complex background,and motion blur.Therefore,it is of great significance to study high-performance video-based human pose estimation methods.Traditional video human pose estimation only uses bone point position loss(such as L2 loss)as the objective function in the network optimization process,and does not consider temporal correlation and mutual constraints between human bone points.Aiming at the inaccurate estimation of skeleton points in traditional human pose estimation algorithms,this paper proposes the motion consistency and structural loss function of skeleton points,and combines the loss of skeleton point position to improve the prediction accuracy of skeleton points.For the estimated interference problem,a temporal attention module is introduced to enhance the saliency of temporal prediction features.At the same time,in view of the multi-person situation in the image and the flexible and changeable posture of the human body,this paper proposes an adaptive Gaussian heatmap,which is combined with the pedestrian features predicted by the network and sent to the network together to improve the robustness of the network prediction.and accuracy.The specific steps are:(1)In the network training stage,the ConvNet1 module is used to obtain the initial predicted value of the skeleton points of consecutive T frames,combined with the true value information of the skeleton points,the prediction deviation in the time series is calculated according to the proposed time series consistency loss function(That is,timing consistency loss),and calculate the structural loss and skeleton point position loss for the output of each stage of the network,and add the losses of each stage to obtain the overall structural loss and skeleton point position loss.Finally,the structural loss,timing consistency loss and bone point position loss are combined to jointly optimize the network.(2)Introduce an Attention Module(AM)in each stage of the network.At each stage,the LSTM module is used to obtain the timing features of the first few frames of images.In this paper,after the intermediate feature map passes the LSTM module,the timing attention module is added,and the output latent state of the LSTM is sent to the attention module to improve the area of interest in space and channels.Feature weights to enhance the saliency of time series prediction features.(3)Use the existing pedestrian detector to obtain the pedestrian detection frame,and calculate the corresponding pedestrian adaptive Gaussian Heatmap(AGH)according to the pedestrian detection frame information,and use the adaptive Gaussian heatmap and the network to predict pedestrian features.Combined with feeding into the network,the prediction accuracy of the network is further improved.Experiments are carried out on the standard datasets Penn Action and Sub-JHMDB to validate our method.The accuracy of PCK keypoint recognition on Penn Action and Sub-JHMDB datasets in this paper reaches 99.1%and 95.3%,respectively.On the Penn Action dataset,it achieves a 1.4%improvement over the LSTM Pose Machines(LPM)model.On the Sub-JHMDB dataset,it is 1.7%higher than the LPM method,and the experimental results demonstrate the effectiveness of the model in this paper.Combined with the video human pose estimation algorithm proposed in this paper,this paper designs and develops a set of motion simulation fitness assistance system based on skeleton point information,which mainly consists of video human pose estimation module,skeleton point-based action recognition module,calorie calculation module,and standard fitness example.It is composed of a display module.The system obtains the exercise images of the exercisers through the web camera,and displays the action recognition results,standard videos and calorie consumption on the GUI interface,so as to help the exercisers to exercise scientifically and effectively.