Research On Train Door And Window Detection Based On Deep Learning

Railway transportation is one of the world's most important means of transportation at present,how to improve the safety of related industries is a long-standing problem.In the process of train safety operation,the vehicle condition monitoring is an important but complicated task.The identification and detection of train doors and windows is the basis for other related research.At present,it can only be checked manually by installing inductive equipment in the doors and windows of trains.However,there are many subjective factors in the manual verification process,which may cause serious hidden dangers to the safety of train operation,such as: human fatigue,distraction,and various false detection.In order to eliminate the influence of human subjectiveness,fatigue and other unstable factors,it is a general trend to have the real-time monitoring of train status in the future to replace manual operation with artificial intelligence.Based on this research goal,this paper proposes a detection method based on the convolutional neural network with variable windows.In this method,a region of interest(Region of Interest,RoI)selection method based on sliding window is designed,which collects the image group with the same center point and multiple receptive field through the gradual window.The elements in the image group are mapped to the same dimensional space through different fully connected layers,and finally the target detection is performed through the basic CNN.At the same time,this paper optimizes and improves the performance,and further proposes a CNN model based on combined multi-layer classifier,which controls the feature extraction of different levels in CNN through sliding window and multi-branch combined multi-layer classifier to improve detection accuracy.Finally,through the experimental analysis and comparison with other advanced methods,the multi-classifier-based detection methods proposed in this paper have achieved high accuracy,which indicates that it can effectively complete the detection and identification of train doors and windows.At the same time,the test results show that,based on the training set,validation set and test set provided in this paper,the detection accuracy and performance of the proposed methods are better than other advanced methods in the task of train door and window safety detection and recognition.