Monocular Human Motion Capture Based On Deep Learning

In recent years,monocular human motion capture is one of the hot research topics in the field of visual algorithms,and it has a wide range of application prospects in the fields of human-computer interaction,virtual reality,animation,and games.This paper mainly studies the problem of human motion capture in monocular video.The problem can be divided into two types: estimation of two-dimensional human body key point coordinates from image space,use of two-dimensional position coordinates to estimate three-dimensional depth information,and use of human body model for body reconstruction.step.Although the current monocular human motion capture technology based on deep learning technology reduces the cost and use scene restrictions to a certain extent compared with traditional motion capture technology,it still takes a long time to calculate,and the reconstructed human motion feet have ground and slippery feet.The problem of walking and floating is not convenient for use in scenes such as film and television animation production and virtual live broadcast.In response to the above problems,this paper proposes a monocular video human motion capture method based on deep learning.The main research work and results include the following:(1)Through the comprehensive research and analysis of the current two-dimensional human body pose estimation algorithm and lightweight target detection network,Based on the design ideas of lightweight target detection network,the Lite Human Pose Net method is proposed.This method has the characteristics of less parameters,fast inference speed,higher estimation accuracy and convenient training.The effectiveness of the proposed method is shown through ablation experiments.(2)Through the research and analysis of the traditional human motion capture system on the foot motion simulation processing method,combined with the current mature deep learning method,a foot touch discriminator combined with the twodimensional posture algorithm is proposed to judge the person in the video frame The condition where the feet touch the ground.The foot contact discriminator can achieve functions similar to wearable sensors and has a lower cost.(3)The shape reconstruction mainly uses the human body model to fit the twodimensional human body coordinates in the image space and the estimated depth coordinates to obtain the reconstructed human body movements.The current depth coordinate estimation technology is relatively mature,and the existing technology is selected.Aiming at the problems of foot penetration,sliding and floating in the reconstructed action,a method for optimizing and fitting the foot problems of the human body model according to the foot contact condition is proposed,and the effectiveness of the proposed method is proved through experimental comparison.Experiments have shown that the monocular video motion capture method proposed in this paper is more suitable for the needs of application fields such as film and television animation production and virtual anchors than the previous method,and the running speed has been optimized to a certain extent.