Research On Climbing Behavior Recognition Based On Human Skeleton

With the acceleration of urbanization and the increase of high-rise buildings,climbing behavior has become a common social problem.Many people ignore the danger of climbing while pursuing excitement and taking pictures.People are increasingly using surveillance videos in their daily lives to ensure public safety.Traditional monitoring technology typically uses video surveillance equipment and security personnel.Security personnel need to stare at the monitor for a long time,which can cause fatigue and neglect,and requires a large amount of manpower costs.At the same time,security personnel need to monitor the screen in real-time and timely detect and handle abnormal situations.However,in large-scale scenarios,real-time performance is difficult to guarantee,often requiring a lot of time and resources.Human action recognition is an important task in computer vision,which has broad applications in the real world.Compared to traditional image-based behavior recognition methods,skeleton-based behavior recognition methods do not need to consider factors such as the subject’s clothing and environmental occlusion,and therefore have stronger robustness.At the same time,the skeleton is composed of the key bones of the human body and can more accurately reflect the posture and movements of the human body.Therefore,skeleton-based behavior recognition methods usually have higher recognition accuracy.Currently,research on skeleton-based behavior recognition has received great attention from researchers.Therefore,this paper mainly focuses on the research of climbing behavior recognition based on human skeleton data,including the following aspects:(1)For human skeleton extraction,this paper first uses the human pose estimation algorithm Open Pose to extract the skeleton data of each frame of the video and obtain the human pose estimation results.To format the Open Pose output for the behavior recognition model,this paper integrates and parses it to generate human skeleton sequence data.The generated skeleton sequence may contain noisy data and inaccurate bone data due to interference from complex situations,which can affect the accuracy and reliability of subsequent recognition.Therefore,this paper filters all humans in each frame based on joint confidence to select clearer,more reliable,and accurate bone data,which is beneficial for subsequent behavior recognition.(2)For the purpose of climbing behavior recognition based on human skeleton data,this paper proposes an improved MST-GCN-based climbing behavior recognition method.Firstly,to solve the problem of a large amount of feature information being continuously weakened during the transmission process in the feature extraction network,this paper introduces the idea of dense connection network into the original spatiotemporal graph convolutional network of MST-GCN,reconstructs the feature extraction network,and obtains richer contextual relationships between joints.Then,to address the problem of insufficient extraction of spatiotemporal and channel information leading to inaccurate action recognition,the convolutional block attention module is introduced into the improved spatiotemporal graph convolutional network,which generates attention feature maps in sequence on the channel and spatial dimensions,and enhances the feature extraction ability of the model on key information.(3)Finally,the proposed climbing behavior recognition method based on improved MSTGCN was trained and tested using the Climb-skeleton dataset constructed in this paper.Through ablation and comparative experiments,the effectiveness of the model was verified.The proposed method was used to recognize the climbing behavior in the test videos of Climb-skeleton dataset,and the recognition results were compared with those of ST-GCN.The algorithm proposed in this paper can achieve more accurate results.