基于重时空注意力的地铁司机动作识别

Subway Driver's Action Recognition Based on Re-spatiotemporal Attention

地铁司机的动作有重复率高、相似度大等特点,这易导致注意力崩溃。因而直接加深网络深度来提升准确率的方法会失效。为了解决深度时空注意力崩溃的问题,提出一种重时空注意力模型,该模型是在Re-attention机制和时空注意力的基础上改进而来。它能很好地解决网络因深度带来的性能饱和等问题。实验表明,重时空注意力在同深度的注意力网络上性能得到了5.5%的提升且不增加额外的计算量,解决了上述存在的问题。

The actions of subway drivers have the characteristics of high repetition rate and high similarity,which can easily lead to attention collapse.Thus,the method of directly deepening the network depth to improve the accuracy will fail.In order to solve the problem of deep spatiotemporal attention collapse,a heavy spatiotemporal attention model is proposed,which is improved on the basis of re-attention mechanism and spatiotemporal attention.It can well solve the problems of network performance saturation caused by depth.Experiments show that the performance of spatiotemporal attention on the attention network with the same depth is improved by 5.5%without additional computation,which solves the above problems.