Research On Foreign Object Detection In Risk Area Between Subway Platform Door And Train Door Based On Top-mounted Visual Sensor

The presence of foreign objects in the risk area between the subway platform door and the train door will pose a threat to the safe running of the train.However,the currently applied nonvisual equipment such as laser sensors has large detection blind spots,low detection accuracy,and cannot visualize the detection results,which can no longer meet the requirements for subway safe operation.Therefore,this paper adopts the foreign object detection scheme based on the top-mounted visual sensor to monitor the risk area between the subway platform door and the train door without blind spots,and uses a variety of deep learning algorithms to gradually improve the foreign object detection accuracy,to assist the subway staff to quickly and accurately Discover hidden safety hazards that are difficult to detect during manual inspections.The main research contents are as follows:First,to achieve high-precision localization and classification of foreign objects,this paper adds channel and spatial attention modules to the convolution module of the backbone network of YOLOv4,so that the network pays more attention to areas with rich foreign object information in the two dimensions of channel and space,so the more critical features are extracted for foreign objects localization and classification.The improved YOLOv4 based on the attention mechanism can improve the recognition effect of small-sized foreign objects,and the performance is more stable under light interference.The m AP@0.5 for evaluating the effect of foreign object localization and classification reaches 94.8%,which is 1.1% higher than that of the YOLOv4 algorithm,and the inference speed reaches 82 frames per second on the RTX2080 Ti,which can meet the real-time requirements that FPS is greater than 30 frames per second.Second,to alleviate the false detection problem of the improved YOLOv4 in non-risk areas such as windows and realize the classification of risk areas,this paper proposes a multi-task deep learning model to simultaneously complete foreign object detection,risk area segmentation and area line detection.The shared encoder can extract deep features and rich semantic information between related tasks,which not only makes the m Io U of risk area segmentation and area line detection reach 96.3% and 94.3%,respectively,but also makes the false detection probability in non-risk areas is reduced by 30.4% relative to the improved YOLOv4.The risk area is classified by sliding search,and it can also be judged whether the foreign object exists in the gap between the train door,the platform door,or the gap between the two doors,so that the subway staff can deal with the foreign objects more quickly and properly.In addition,the multi-task model uses lightweight modules,and the inference speed reaches 45.3 frames per second on the RTX 2080 Ti.Third,to filter out the prediction results that do not meet the scene characteristics and further improve the accuracy of foreign object detection,this paper integrates the risk area segmentation,area line detection and foreign object detection results of the multi-task model,making full use of the scene properties that the foreign objects exist in the risk area and break area lines.Compared with the original multi-task model,the fused multi-task model reduces the probability of false detection in non-risk areas by 67%,and improves m AP@0.5 by 4.3%on all test sets.Therefore,using the fused multi-task model for foreign object detection can assist subway workers in more accurately identifying potential safety hazards.