Gesture is one of the most important body language of human beings.Gestures account for the largest proportion of all human gestures.Gesture estimation is the use of computer technology to analyze and estimate the posture of human hand movements.In the current intelligent age,gesture estimation is crucial in the application of human-computer interaction,and has great research significance and value.Gesture estimation based on depth images has achieved good results in recent years,mainly due to the emergence of sophisticated data acquisition equipment depth cameras and deep learning based algorithms.However,depth image-based gesture estimation is very limited in practical applications,so color image-based gesture estimation has begun to appear in the past two years.However,due to some characteristics of the hand itself,such as high flexibility,low resolution and difficulty in key point recognition,the most important thing is to estimate the 3D pose from 2D pictures without depth information at all.These challenges have made the current color image based gesture estimation less than ideal.In order to solve some problems encountered in color image gesture estimation,this paper proposes the following innovative solutions:(1)Using Hourglass Network for hand key detection,due to the network characteristics of Hourglass Network,it can move the positional relationship between the key points of the handshake from the whole to the local.(2)The SSD detection algorithm is used to locate the position of the hand in the picture,and the 2D estimation process of the key point and the hand positioning process are merged.In this way,when the hand positioning is performed,the detection of the 2D key points can be supervised,and the position of the hand is positioned by the segmentation method,thereby avoiding the continuous pixel classification,and the timeliness is greatly improved.(3)The feature maps of different levels and different scales are merged.Predicting on different levels of feature maps to adapt to key points of different difficulty levels,and combining top-down and bottom-up features at various levels,so that the estimation takes into account the semantics.The feature also does not neglect the detail features and is effective for low resolution gesture estimation.In view of the above innovative methods,this paper designs a gesture estimation algorithm based on Hourglass Network and a gesture estimation algorithm based on multi-level feature fusion,and carries out related experiments on gesture data.The results show that our method is effective and Some details have more outstanding performance. |