Object detection at level crossings using deep learning techniques

摘要

Level crossings pose a significant risk, and their use or misuse can lead to serious accidents between railway vehicles and road and footpath users who may sometimes behave unpredictably. Currently, Great Britain has 7,500 level crossings and the Rail Safety and Standards Board (RSSB) reported 3 fatalities and 385 near misses in the year 2015-16. The Office of Rail and Road (ORR) and RSSB have recommended the upgrading of the automated technologies used at level crossings. For this purpose, automated safety systems remain a key area for investigation for inclusion in interlocking systems at level crossings. In this study, we propose a solution to automate the operational cycle at level crossings using deep learning technologies. The proposed solution will add another layer of resilience to the safety system without the use of manual operators at level crossings and ensure an interlocking system with a safety integrity level (SIL) of 3-4. The paper discusses the current technologies utilised at level crossings along with the algorithms required for post-processing the data. The proposed model integrates deep learning technology with the current vision system, CCTV, to detect and localise an object at a level crossing. Two approaches are discussed to train the neural network: from scratch or using transfer learning techniques. Neural networks along with their associated accuracy, which represent the percentage of true predictions on the test dataset is mentioned as well. Classification, detection and segmentation models are used to classify, detect and localise objects like vehicles, bicycles and pedestrians at level crossings. Neural networks trained from scratch achieved an accuracy of 75%, networks trained for object detection using transfer learning techniques achieved an accuracy from 58 to 82%, and the image segmentation model, Mask RCNN, achieved an accuracy of 95% on test data. Finally, this paper discusses some future work to improve the network's accuracy and another application of convolutional neural networks (CNN) using radar sensing.