A Detection Methods With Image Recognition for Specific Obstacles in the Urban Rail Area

As the automation level of urban rail transit is becoming higher, the safer operation of rail transportation systems is playing a crucial role in ensuring the lives and property of passengers. However, the external environment of rail transit is complex and dynamic, especial the various foreign object intrusions, which severely threaten the safety of urban rail. This study proposes a novel obstacle detection method for rail track areas by integrating 2D and 3D object detection techniques. This method employs a two-branch deep neural network that extracts multi-scale texture features in the 2D image branch while simultaneously learning the spatial structure features of targets in the 3D image branch. Then, the backbone networks of the two branches are fused through a feature fusion module. Network pruning reduces network computation by 39% while reducing mAP by only 0.5 percentage points. Finally, the experimental results demonstrate that the detection methods with image recognition for specific obstacles achieves high detection accuracy in different environments and detection distances. Under the typical detection distance of 90m, the pedestrian detection accuracy mAP value reaches 91.2%, the distance measurement error MAE value is 0.96m, and the frame rate is about 25 FPS.