Action Recognition System Based On Improved LM Skeleton Fitting Algorithm

With the continuous development of artificial intelligence,one of its important applications in the field of computer vision,human action recognition technology is gaining more and more attention.Human action recognition technology mainly uses computer vision to determine the category of human action behavior in a picture or a video,and the technology can be applied in many fields such as human-computer interaction,medical rehabilitation,and motion monitoring.The existing research methods based on computer vision are easily affected by various factors such as complex background,human wearing,and light intensity in the process of human body recognition,resulting in a low recognition rate and poor recognition results.To address the above problems,this paper proposes a method to recognize human actions by predicting human skeletal points,which is mainly implemented by improving the 3D human skeletal points of Kinect and OpenPose.Firstly,the traditional image sensor is replaced by the Kinect depth sensor,which can conveniently acquire human color image data,depth image data,and skeletal point data.Secondly,the skeletal points are processed and optimized by combining the OpenPose 2D skeletal point detection algorithm,which is an RGB image-based algorithm model that can detect and recognize single OpenPose is an algorithm model based on RGB images,which can detect and recognize 2D skeletal point data of a single person or multiple people,and can infer and predict the location of the occluded skeletal points in the case of a small part being occluded.Finally,the 3D skeleton point coordinates obtained from Kinect are used as the coordinates to be optimized,and the 2D skeleton point coordinates obtained from OpenPose are mapped and processed to obtain the 3D skeleton point coordinates as the optimization parameters,and the LM algorithm is used to fit and optimize both of them to obtain the required stable 3D human skeleton point sequence.At the same time,to address the problems of large operation and slow speed of the LM algorithm,the process of the LM algorithm is improved by adjusting the constraint coefficients and constraint terms,which reduces the operation and improves the recognition speed.Through quantitative testing and qualitative analysis,the improved human action recognition algorithm is slower than the recognition speed of Kinect or OpenPose algorithm alone by about 0.6-1 second,but the skeleton points obtained are more stable and effectively solve the problems of jittering and self-obscuring of skeleton points during Kinect recognition,and at the same time,when combined with DTW algorithm for human action recognition evaluation,the average The recognition rate is 6.32% higher than that of Kinect alone,and the improvement of the recognition rate of self-obscuring action is especially obvious.