Rail image recognition technology based on deep learning

In order to ensure the operation safety of the subway in urban rail transportation and avoid safety accidents caused by obstacles inside the railway tracks, it is necessary to conduct subway rail recognition. Considering that the railway tracks have slender and continuous physical characteristics, and inspired by the lane detection in deep learning, we proposed the CLRNet-L algorithm for railway track recognition, which is an improvement from the CLRNet algorithm. To solve the problem of long and thin railway tracks that are difficult to be accurately identified and localized, CLRNet-L used a feature pyramid network to extract and fuse high-level features and low-level features. Through the idea of top-down, the high-level features were first used to locate the railway tracks in a rough way, and then the shallow features fused with the high-level features were used to further refine the tracks so as to realize the identification of the railway tracks. In response to the problem of railway tracks that are difficult to distinguish from the surrounding environment due to their dark colors, attention mechanisms and multi-scale aggregators were introduced into the original CLRNet. We proposed a large kernel attention module to capture more contextual information and enhance the feature representation of railway tracks. Due to the lack of public railway track datasets for rail recognition in the field of rail transit, we used track photos collected during subway operation in Hangzhou Line 5 and Line 6 to create a dataset of track scenes. The dataset, which includes rail images of straight, curved, and turnout scenes, was used for rail recognition experiments to verify the effectiveness of the algorithm. Experimental results show that CLRNet-L achieved 88.96% MIoU and the fastest detection speed of 11.54 ms in the custom dataset, which has higher accuracy and detection speed compared with other detection algorithms. The research results provide a technical foundation for subway safety technology, especially obstacle detection, to ensure the safety of subway operations. © 2024, Central South University Press. All rights reserved.