Research And Application Of Human Skeleton Extraction And Action Detection Segmentation Based On Deep Learning

Based on the development of computer technology,giving the visual capabilities to computers is an important guarantee for achieving noncontact intelligent interaction between the computer and people.As the main role of social activities,human actions convey critical information.Therefore,the use of artificial intelligence technology to identify human actions in videos has important academic significance and application value.However,the complexity of the background easily affects the results of motion detection,so the realization of action detection tasks based on human skeleton information has shown great advantages.Under this background,this study firstly studied the problem of human body pose estimation,using top-down method and bottom-up method to achieve the location of human skeleton points.Among them,the top-down method uses a network model based on LSTM units,and an average PCK accuracy of 93.72% is obtained on the J-HMDB dataset.The bottom-up method uses the Res Net50 network model based on dilated convolution and Part Affinity Fields,and achieves an average prediction accuracy of 85.9%on MPII dataset.Afterwards,this paper studied the task of Temporal Action Detection,using the R-C3 D network to achieve the positioning and classification of human action.On the THMOUS2014 dataset,the R-C3 D model was verified based on the original video information and the human skeleton information,and the accuracy of 21.8% m AP@0.5 and 26.6% m AP@0.5was obtained separately.Among all the 20 types of actions in the dataset,the accuracy of prediction of 16 types has been improved through pose estimation processing,which verifying the advantage of using skeleton information as input of Temporal Action Detection.Considering the application value of human action detection and segmentation,this paper has studied the automatic assessment of workers' operations.First,a dataset was established based on a certain factory.Afterwards,the top-down method and the bottom-up method were used to predict the skeleton of the self-collected data,and the accuracy of 71.02%and 84.06% was obtained separately.Further,the original video information and skeleton information are used as the input of the sequential motion detection model to segment and classify the defined motions in the long video,and the accuracy of 84.3% m AP@0.5 and 85.9% m AP@0.5 is obtained separately,The advantage of using skeleton information as input to complete sequential motion detection tasks is verified again.As a comparison,the above two motion prediction results are fused and an accuracy of 87.8% m AP@0.5 was obtained.Finally,according to the experimental results in this study,some feasible optimization directions are proposed for human pose estimation and time-series motion detection.