基于骨架数据的列车司机异常行为检测

Detection of Abnormal Behavior of Train Drivers Based on Skeleton Data

摘要: 针对列车司机异常行为检测时存在准确率以及效率低的问题，在时空图卷积网络(Spatial Temporal Graph Convolutional Network, ST-GCN)行为检测模型的基础上，提出一种基于骨架数据的列车司机异常行为检测(Spatiotemporal Graph Attention and Multi-Time Scale Temporal Convolutional Network, ST-GAT)模型。通过利用骨架数据进行建模，并引入图注意力网络(Graph Attention Network, GAT)模块，利用动态关注机制以提升ST-GCN模型对空间特征的提取能力。通过提出的多时间尺度时域卷积网络(Multi-Scale Temporal Convolutional Network，MS-TCN)模块进行不同时间尺度下的时序提取特征，扩大时域卷积网络(Temporal Convolutional Network, TCN)模块感受野范围，克服原始TCN模块不灵活、无法检测不同时间尺度的问题，提高了模型的准确率，同时使用交叉熵损失函数克服了模型训练过程收敛速度慢的问题。实验结果表明，本文方法在测试集上比ST-GCN模型准确率提高8.8%，FLOPS提高2%。因此，所提出的方法在提高列车司机异常行为检测的准确率和效率方面表现出较好的性能。

Abstract: Aiming at the challenges of low detection accuracy and efficiency in train driver abnormal behavior detection, this study proposes a model based on skeleton data, named Spatiotemporal Graph Attention and Multi-Time Scale Temporal Convolutional Network (ST-GAT), built upon the Spatial Temporal Graph Convolutional Network (ST-GCN) behavior detection model. By leveraging skeleton data for modeling, the model introduces a Graph Attention Network (GAT) module with a dynamic attention mechanism to enhance the ST-GCN model’s extraction of spatial features. Additionally, the model incorporates a Multi-Scale Temporal Convolutional Network (MS-TCN) module to extract temporal features at different scales, addressing the flexibility and scale limitations of the Temporal Convolutional Network (TCN) module in ST-GCN. To improve training efficiency, the model adopts the cross-entropy loss function to expedite convergence. Experimental results demonstrate an 8.8% improvement in accuracy and a 2% increase in FLOPS on the test set compared to the ST-GCN model. Therefore, the proposed method exhibits favorable performance in enhancing both the accuracy and efficiency of abnormal behavior detection.