| Human posture estimation has broad application prospects in human-computer interaction,motion capture,intelligent monitoring,and other fields.It is essential for machines to effectively understand and analyze human actions for intellectual life.Up to now,the detection accuracy of the human posture estimation algorithm on the standard data set has reached a reasonably high level,but how to reduce the parameter amount and calculation amount of the model while maintaining high accuracy for deployment on embedded devices is one of the critical contents of the current research.Based on this,this paper designs a light weight human posture estimation model,which has the advantages of low parameter number,low computation,high accuracy,and fast convergence speed.In the backbone network part of the model,an improved light weight feature extraction network is proposed,which can better extract features while maintaining the lightweight of the model;In the position-coding part of the model encoder,a learning Fourier position coding method is proposed,which can solve the problems such as the lack of edge position information of traditional position coding and the lack of flexibility of predefined wavelengths,and improve the accuracy of model detection.The specific research contents are as follows:(1)A human posture estimation model based on the combination of light weight convolutional network and Transformer encoder is proposed.This model first designs a light feature extraction backbone network of depth separable convolution plus depth residual block,which can effectively extract features and fuse feature information of different scales.Only a few network layers and parameters are used to achieve a good feature extraction effect.At the same time,an optimized Transformer encoder is designed to capture the spatial position relationship between key points,effectively solving the problem of poor high-dimensional feature extraction ability of lightweight convolutional network,and enabling the model to obtain global constraints from a higher resolution to find the spatial relationship between key points,making key point positioning more accurate.The validity of the proposed model is proved by comparative experiments.(2)A learning Fourier feature position coding method is proposed.This coding method uses the parameters in the Fourier function to represent the relationship between the spatial positions of key points.The translation invariance of the Fourier function can effectively avoid the problem that the traditional sinusoidal position coding has a weak perception of the relationship between the edge feature spaces,and strengthen the ability of Transformer encoder to capture the various positions of human key points.Using the invariable parameter quantity of the learnable Fourier feature,the feature position information is converted into a fixed and learnable parameter and learned synchronously with the model training process,which solves the problem of parameter explosion of commonly used learnable position coding parameters with the increase of the model sequence.The encoding method of feature position is optimized from two aspects,which further improves the detection accuracy of the model. |
