Deep Learning-based Algorithms For Visual Perception Of Railroad Environment

Full automatic operation has become the new development direction of rail transportation.The problem at present,however,is that most fully automated driving lines operated under human supervision.The solution is to establish an on-board vision system to intelligently monitor and effectively perceive the train’s operating environment in real-time.Existing methods mainly involve installing video monitoring equipment at the trackside or regular manual inspections,which are inefficient,costly,and cannot meet the long-distance and all-weather environmental monitoring of the railways.The establishment of an environment perception system based on an onboard vision for trains is important for fully automated operation,which is the way forward for intelligent railways.Based on deep learning technology,multiple visual problems in the perception system of fully automatic trains are discussed in this dissertation.In conjunction with railroad feature analysis,the research focuses on common problems in studying low-level feature extraction and establishing the ability to perceive the environment from scene and object perspectives.To improve the adaptability of visual algorithms to complex light sources in the railway environment,an image enhancement algorithm was used to solve the exposure problem.This dissertation focuses on exploring and solving several key technical problems of visual environment perception using deep learning algorithms,and the main contents and innovations of this dissertation are as follows.Firstly,the edge detection algorithm based on low-level visual features is researched.Edge detection and recognition are prerequisites for many computer vision tasks.The structure of deep neural network,data and loss function will affect the low-level feature extraction in multiple visual tasks.In this dissertation,a multi-scale feature hybrid algorithm based on dense residual connections is proposed for edge detection.The algorithm performs feature extraction and high-level processing in the encoder-decoder structure.Feature enhancement unit and residual up-sampling block are designed to realize the effective use of features in the decoding network.By analyzing the intrinsic causes of the class imbalance problem,a biased cross-entropy loss function is proposed for reducing the training risk of neural networks under class imbalance data.In the experiment,an ablation study was set up to verify the functions of related modules.Experimental results on several data sets prove the rationality and advancement of the algorithm in network structure design and training methods.Secondly,the accurate identification of railroad areas is researched from the perspective of scene segmentation.The railroad area can be identified as a closed scene component in the image.Based on this intuition,a fully convolutional railroad segmentation network called Rail Net is proposed.It differs significantly from the traditional method of determining the railroad area by detecting two railroad lines.Rail Net combines multi-scale feature extraction and pyramid feature up-sampling to incorporate different levels of features into the backbone network.A V-shaped fully convolutional segmentation network is designed for segmentation tasks to restore feature resolution layer by layer.To train the proposed railroad segmentation network,the RSDS dataset and transfer learning are used to update the network parameters.The experimental results show that the algorithm proposed in this dissertation can segment the railroad region stably and accurately under different scenarios.Thirdly,the small target pedestrian detection algorithm in the railroad region is researched.By analyzing the problems of current object detection algorithms and the performance requirements of onboard vision algorithms,a single-stage pedestrian detection algorithm is proposed using a combination of telephoto and wide-angle cameras as the vision hardware platform.The algorithm establishes a backward feature enhancement channel in the basic feature extraction network and builds a convolutional detector based on the enhancement features to perform object detection.The new prior detection box strategy is used to effectively reduce the number of detection boxes and enhance the detection ability of the small object.Hard example mining and the focal loss strategy are used in the network training to solve the positive and negative example imbalance problem.Multiple experiments are designed to demonstrate the effectiveness of the proposed algorithm.Finally,the method of improving the image quality of onboard vision under variable and complex exposure conditions is researched.By analyzing the exposure of onboard video images in the railroad environment,a real-time enhancement algorithm based on self-supervised learning in the low-light environment is proposed.The algorithm uses the densely connected structure as the backbone network to establish a feature-size invariant network to extract the image illumination,color,edge and other information and output the enhancement rate map.The non-linear mapping function is used to achieve high dynamic range pixel illumination adjustment.Through the hierarchical structure,the exposure rate of the low-light image will continuously be enhanced from low-level to high-level.The algorithm uses a self-supervised method to train network parameters and uses the low-light image’s characteristics and prior knowledge to construct a loss function,thus solving the problem of the difficulty of obtaining a large amount of labeled training data in the railroad environment.Experimental results of image enhancement in various scenes show that the proposed method can adapt to the exposure values of the input image.The algorithm dynamically adjusts the exposure rates of low and high exposure areas to improve the image quality of low-light images.