基于改进SlowFast算法的电梯乘客异常行为识别

Abnormal behavior recognition of elevator passengers based on improved SlowFast networks

目的:解决轿厢式电梯内乘客异常行为识别算法无法有效利用视频数据中时序特征的问题。方法:提出一种基于残差支路与BiFormer改进的SlowFast网络算法。该网络结构以RGB视频帧和残差帧作为输入,以多支路提取特征信息,融合慢支路、快支路和残差支路的时空特征,增强对乘客异常行为的敏感性,降低背景变化带来的影响。为增强时间维度信息的有效利用,在快支路与残差支路引入BiFormer结构,以学习帧间关联信息,从而提高网络对乘客异常行为识别准确率。结果:为验证网络算法的有效性,以电梯乘客异常行为数据集验证提出的网络结构。与原SlowFast网络进行对比,改进后的网络识别准确率提高了8.46%。结论:结果表明,所提出的网络算法能够充分利用视频帧中时间维度信息,可有效提高电梯乘客异常行为识别准确率,且在电梯内背景与光线变化较大的情况下,仍然具有较好的识别效果。

Aims:This paper aims to solve the problem of the ineffective utilization of temporal features in video data of the algorithm for identifying passengers'abnormal behavior in elevator cabins.Methods:A SlowFast network algorithm based on residual branches and BiFormer improvement was proposed.This network structure took RGB video frames and residual frames as inputs,extracted feature information from multiple branches,integrated the spatiotemporal features of slow branches,fast branches,and residual branches to enhance the sensitivity to passenger abnormal behavior,and reduced the impact of background changes.Results:To verify the effectiveness of the network algorithm,a dataset of abnormal behavior of elevator passengers was used to validate the proposed network structure.Compared with the original SlowFast network,the improved network increased the recognition accuracy by 8.46%.Conclusions:The results showed that the proposed network algorithm could fully utilize the temporal dimension information in video frames and effectively improve the accuracy of identifying the abnormal behavior of elevator passengers.It has good recognition capability even inside elevator cabins with different background and illumination.