MRSA-Net: Anomaly Detection of Metro Stations Based on Residual Attention Network With Path Aggregation and Efficient Representation

Anomalous objects in the gaps between train doors and platform screen doors may threaten passenger safety or even cause serious accidents, but anomaly detection in this risk space is rarely explored. Anomaly detection aims to simultaneously predict the category and localization of anomalies. In this article, we propose a novel metro risk space anomaly detection network (MRSA-Net) consisting of three parts: residual attention module, involution-based path aggregation module, and generalized focal loss (GFocal) head module. The residual attention module is an improved attention mechanism that combines transformer attention with only key content factors and modulated deformable convolution. It significantly reduces floating-point operations (FLOPs) by sharing the computational burden between these two components. The involution-based path aggregation module is designed to address the issue of visual element loss in different spatial positions during information propagation. The GFocal head module includes two representations: the classification–intersection-over-union (IoU) joint representation and the general distribution representation. The former is introduced to address the issue of inconsistent usage of classification scores and IoU scores during training and inference, while the latter is introduced to handle anomalous objects with blurry manual labels caused by insufficient lighting conditions. We conduct extensive experiments on the metro risk space anomaly (MRSA) dataset and the large-scale COCO dataset, and the average precision (AP) can reach 86.8% and 43.8%, respectively, which outperforms previous methods.