Human Behavior Recognition Based On ST-GCN And Skeleton Information

Because bone-based action recognition is not affected by the physical characteristics of the human body and its potential advantages,that it can simply and clearly convey the important information of human behavior recognition,which makes it a hot topic that has been discussed by many former researchers in the field of computer vision.However,traditional skeleton modeling usually relies on the artificial setting of traversal rules,which leads to limited expression ability and difficulty in promotion.Some methods do not pay attention to the movement connection between body parts in the action recognition process,which leads to the recognition efficiency that fails to meet expectations.In response to these problems,this article proposes different improvement methods based on ST-GCN.The main contents of the work are as follows:(1)Based on the ST-GCN model,a new partitioning strategy for dividing the skeleton joint points is proposed: First,use Open Pose to extract the skeleton joint points,construct a time-space map in the skeleton sequence,and then strengthen the relationship between the relative positions of the body.The relationship constructs a new division strategy to improve the association of skeleton joint point information in time and space.Then use the frame sequence spatial information to extract the skeleton information to generate a higher-level feature map,and finally use the standard Soft Max classifier for action classification and recognition.(2)A novel training strategy is proposed to improve the convergence of the training model.The adaptive learning rate is adjusted by using the linear rule of hyper-parameters,and the model is trained according to the method of gradual attenuation of the learning rate.The simulation results show that while the stability of the training model's convergence is effectively improved,the training time will not increase.(3)Based on the ST-GCN model,a new method of processing video data and skeleton joint diagrams is proposed,which can effectively solve the interference of ambiguous frames caused by the complex environment background and shooting methods.First,use Markov Decision Process(MDP)for video data to focus on effective information frames.Secondly,the dependency relationship between the joints in the skeleton joint diagram of the selected key information frame is captured,so as to strengthen the spatial information association of the ST-GCN model.(4)Verify the two ST-GCN-based models proposed in this article on the two public data sets Deepmind Kinetics and NTU-RGB+D(divided into X-Sub and X-View two data sets)for human behavior recognition Method of improvement.According to the experimental results,the partition strategy that combines the strengthening of the relationship between the relative positions of the body and the extraction of skeleton joint point information effectively improves the recognition accuracy of human behavior recognition.However,by extracting key information frames and capturing the dependency between the joints of the image-based convolutional neural network,the problem of poor motion stability of large shots in large-scale data sets can be significantly improved.For the training strategy method proposed in this article,this article has been verified on the two data sets of Kinetics and NTU-RGB+D(X-Sub and X-View).Experiments show that when training in the same environment,the learning rate is adjusted by the training strategy to improve the stability of model convergence,and the top-1 recognition accuracy of the corresponding three data sets has been improved to a certain extent.Compared with the current mainstream methods,the recognition accuracy of the method proposed in this article under large-scale data sets is more competitive.