Dangerous Action Recognition Algorithm Based on Motion Capture

Motion capture relies on computer technology to convert motion trajectories into digital movements, and combines optical equipment, deep learning algorithms and scene modeling to track, measure and record the motion trajectories of human bodies, animals and other objects in real three-dimensional space. Typical application scenarios include industrial high-risk operations, medical rehabilitation monitoring, intelligent security monitoring and sports risk prevention and control. This paper first studies the key techniques and algorithms of motion recognition, and then chooses the most widely used computer vision technology to carry out research. The core work of the research is to build the mathematical model of SIFT algorithm based on Gaussian image and Gaussian difference image, and improve the scale space and introduce MN-SIFT feature descriptor. At last, two accuracy evaluation indexes, top-1 and top-5, were used to conduct experiments on NTU RGB+D 60 dataset, and the experimental results were compared and analyzed.