Research On Human Action Recognition Based On Human Keypoint Detection

Human action recognition is widely used in the age of artificial intelligence,and keypoint detection is playing an important role in the field of action recognition.However,there are still some problems that need to be solved.Firstly,it is difficult to obtain accurate keypoints from a two-dimensional image,therefore this paper proposes an efficient method to solve the problem by using a segmentation neural network.The proposed method can improve the accuracy of keypoint detection.Secondly,it is difficult to find the relationship between human keypoints and various categories of human action,therefore this paper proposes an idea that we can learn the relationship by designing a neural network,which can enhance the accuracy of action classification.The main work and contributions of this paper are as follows:(1)In order to solve the contradictions between detection precision and real-time,this paper proposes an improved Mask RCNN algorithm which could be used to detect human keypoint.The mask branch of the Mask RCNN can generate the high quality prediction of masks,thus improving the accuracy of the human keypoint detection results.The experimental result shows that the proposed method can not only solve the problem of missed detection caused by the top-down method,but also solve the problem of keypoint mismatch caused by the bottom-up method.The experimental results also prove that the proposed method can meet real-time's demands.(2)The human key point detection algorithm has been optimized and improved in three aspects: firstly,in order to solve the problem that the location of Ro I is not accurate,this paper proposes K-means clustering to improve the initialization of the anchor boxes,making it more accurately pinpoint the target areas.Secondly,in order to solve the problem of missed detection caused by multi-person detection,the Gaussian penalty function is used to improve the non-maximum suppression,as a result,improving the detection accuracy of the model.Thirdly,for the problem of the loss of details in the traditional Ro I pooling layer,the Pr Ro I pooling method is introduced to make the model integrate more details.The experimental results show that detection accuracy is well improved by the above methods.(3)In order to solve the problem that the traditional feature extraction methods rely on manual design,we consider to use the fully connected neural network to learn the mapping relationship between the keypoint and action categories.This method can explore deep relationship between the keypoint and the action.In the end,our work implement an action recognition system which uses the human keypoint detection technology.The experimental results show that the proposed method has a high performance than traditional methods and mainstream methods.