The automatic detection of foreign objects between platform screen doors (PSDs) and metro train doors significantly affects personnel and property safety and maintains the train’s normal operation. However, some existing works only determine the presence of foreign objects but cannot indicate their categories. Besides, although deep-learning-based object detection algorithms can indicate the presence and categories of foreign objects, most of them only harness the information in region proposals, ignoring global contextual information. Furthermore, their performance comes at the considerable cost of computational complexity, and leading cannot be well deployed in the metro environment. To address these issues and better implement foreign object detection (FOD), we present You Only Look Once-Transformer (YOLO-Former), a simple but efficient model. YOLO-Former is accomplished based on YOLOv5 through the following procedure. First, the vision transformer (ViT) is introduced for dynamic attention and global modeling, thereby solving the problem that the original YOLOv5 only utilizes information in region proposals and has insufficient ability to capture global information. Second, the convolutional block attention module (CBAM) and Stem module are used to improve feature expression ability further and reduce floating point operations (FLOPs). Finally, we design various variants with different widths and depths to meet every need. Experiments on the foreign object detection dataset (FODD) and PASCAL VOC dataset demonstrate that YOLO-Former-x consistently outperforms other state-of-the-arts with significant margins (0.5 to 11.3 mean average precision, mAP, on FODD and 0.6 to 13.6 on PASCAL VOC dataset). Last but not least, YOLO-Former-x maintains real-time processing speed (27.32 and 28.17 frame per second, FPS, on TITAN Xp).

Platform Screen Doors (PSDs) have been widely used in modern Asian and European metro systems due to the advantages of safety, comfort for passengers. Unfortunately, someone or something will be caught by PSDs and metro doors occasionally, which may lead to serious accidents. Therefore, the foreign object detection between PSDs and metro doors is a burning problem. Moreover, this problem is a challenging and still largely under-explored topic. In recent years, we have seen significant improvements in generic object detection built on deep learning techniques. Accordingly, this paper adopts deep learning technologies to address the problem of foreign object detection between PSDs and metro doors. To the best of our knowledge, this is the first attempt to use deep learning to solve the problem. To realize this, a dataset including 984 real-world images (with 600 × 480 pixels) labeled for six types of foreign objects (bag, bottle, person, plastic bag, umbrella, other) is developed. Then, we compared the performance of some state-of-the-art object detection algorithms (such as You Only Look Once -YOLOv3, Single Shot MultiBox Detector -SSD, and CenterNet) on the dataset. Experimental results demonstrate that the foreign object detection algorithms based on deep neural networks have achieved excellent results, which not only improves the accuracy of detection but also give the categories of foreign objects. YOLOv3 - tiny can achieve the fastest detection speed, up to 200 Frame Per Second (FPS); CenterNet can achieve the best detection results, up to 99.7% mean Average Precision (mAP).

The detection of a fallen person (FPD) is a crucial task in guaranteeing individual safety. Although deep-learning models have shown potential in addressing this challenge, they face several obstacles, such as the inadequate utilization of global contextual information, poor feature extraction, and substantial computational requirements. These limitations have led to low detection accuracy, poor generalization, and slow inference speeds. To overcome these challenges, the present study proposed a new lightweight detection model named Global and Local You-Only-Look-Once Lite (GL-YOLO-Lite), which integrates both global and local contextual information by incorporating transformer and attention modules into the popular object-detection framework YOLOv5. Specifically, a stem module replaced the original inefficient focus module, and rep modules with re-parameterization technology were introduced. Furthermore, a lightweight detection head was developed to reduce the number of redundant channels in the model. Finally, we constructed a large-scale, well-formatted FPD dataset (FPDD). The proposed model employed a binary cross-entropy (BCE) function to calculate the classification and confidence losses. An experimental evaluation of the FPDD and Pascal VOC dataset demonstrated that GL-YOLO-Lite outperformed other state-of-the-art models with significant margins, achieving 2.4–18.9 mean average precision (mAP) on FPDD and 1.8–23.3 on the Pascal VOC dataset. Moreover, GL-YOLO-Lite maintained a real-time processing speed of 56.82 frames per second (FPS) on a Titan Xp and 16.45 FPS on a HiSilicon Kirin 980, demonstrating its effectiveness in real-world scenarios.

Abnormal objects in the risk space between the platform door and train will threaten the safe operation of the metro transportation system. However, selecting the right time to perform anomaly detection during the process of metro in and out station is the primary problem. The trouble of lacking abnormal samples, the high imbalance of abnormal categories, and the defects of abnormal regeneration are not well resolved in most reconstruction-based anomaly detection methods. To handle the aforementioned issues, a method based on light strip inductive key frame extraction and patch-level unsupervised network is proposed. The light strip inductive feature is designed to simulate the status of train door movement, which contributes to extracting key frames for anomaly detection. An unsupervised network, which is named metro anomaly generative adversarial network (MAGAN) and based on dilate fine-grained generator, multipatch discriminator, and local attention reconstruction loss, is proposed for anomaly classification and localization. Our network only utilizes normal samples for training and greatly suppresses the reconstruction of abnormalities by dilate fine-grained generators during testing. The discriminator and local attention reconstruction loss promote the interaction of global and local information in the image and provide a new multipatch-based anomaly localization scheme. The proposed method is tested on the metro risk space anomaly dataset, and the AUCs of anomaly classification and localization achieve 97.2% and 96.4%, respectively, which outperforms those of the state-of-the-art methods.