摘要
洗手动作需关注手部运动时序特征,采用3D-CNN网络或光流法对其进行动态识别时存在计算量较大、实时性较差等问题。为此建立洗手动作数据集,并基于TSM(时移模块)思想提出一种动态手势在线识别的轻型网络TSM-MobileNetV3。首先针对高性能网络MobileNetV3结构特点,设计了一种TSM模块融合方式,构成本网络的残差结构;其次通过对比不同的TSM移位比例、时移长度以及共识融合策略对识别准确率的影响,确定了网络结构与参数。实验表明,在CPU上运行时,此方法在31ms的延迟下可以达到99.872%的识别准确率。
Hand washing actions need to pay attention to the temporal characteristics of hand movement,but the use of dynamic gesture recognition methods based on 3 D-CNN network or optical flow graphs has problems such as a large amount of calculation and poor real-time performance.In this paper,a hand washing action data set is established for this purpose,and a lightweight network TSM-MobileNetV3 for online recognition of dynamic gesture is proposed based on the idea of TSM(Temporal Shift Module).First,according to the structural characteristics of the high-performance network MobileNetV3,a TSM fusion method is proposed to form the residual structure of TSM-MobileNetV3;secondly,the network structure and parameters were determined by comparing the influence of different shift proportion,time shift lengths,and consensus fusion strategies on recognition accuracy.Experiments show that when running on CPU,this method can achieve a recognition accuracy of 99.872% with a delay of 31 ms.
出处
《机电一体化》
2022年第4期3-10,共8页
Mechatronics
