基于改进残差注意力网络的坐姿识别方法

Improved Deep Residual Network Based on Sitting Recognition Method

不良坐姿是导致人类腰椎受损的重要因素,坐姿识别和不良坐姿纠正逐渐成为生命健康领域重要研究方向。传统基于视频图像的坐姿识别算法识别准确率低且占用存储空间大。研究通过薄膜压力传感器采集臀部坐姿压力数据,并利用十字加权滤波法对坐姿压力矩阵进行数据增强与细化,提出一种基于改进深度残差网络的坐姿识别方法。实验结果表明,该方法综合坐姿识别率高达99.84%,该方法结合深度可分离卷积与基于全局的空间注意力机制对残差网络进行改进,提高了网络对局部特征的提取能力,优化了传感器识别坐姿存在的系统失真问题。

Bad sitting posture is an important factor that causes damage to human lumbar spine so that sitting posture recognition and correction of bad sitting posture are becoming important research directions in the field of life and health.Traditional video image based sitting posture recognition algorithms have low recognition accuracy and take up large storage space.In this study,a sitting posture recognition method is proposed based on improved depth residual network by collecting hip sitting pressure data through a thin-film pressure sensor with cross-weighted filtering method to enhance and refine the sitting pressure matrix.The experimental results show that the comprehensive sitting posture recognition rate of this method is as high as 99.84%.The method combines depth-separable convolution with global-based spatial attention mechanism to improve the residual network,which improves the extraction ability of the network for local features and optimizes the systematic distortion problem that exists in the sensor recognition of sitting posture.