Train Active Environment Perception Algorithm Based On Deep Learning

With the continuous expansion of the urban population and the increasing demand for transportation,a series of train rear-end collision accidents put forward higher requirements for operation safety.When the signal system failure causes the train to degrade,the assisted driving system can effectively reduce the occurrence of transportation accidents,while the active environment awareness algorithm is an important part of the train assisted driving system.Traditional train detection and track segmentation algorithms are limited by hand-designed features.Cumbersome steps and complex computation cannot meet the requirements of real-time and accuracy.Furthermore,they are easily affected by environmental factors such as light,weather,terrain,which brings about low recognition rate and wrong segmentation.With the development of deep learning and the advancement of sensor technology,theoretical research has been widely used in many fields,such as PSD anti-pinch recognition,passenger behavior recognition,pedestrian flow detection,etc.But research in the field of train environment perception is still less involved.Aiming at the above problems,this paper conducts train environment perception research based on computer vision algorithms and deep learning theories.Replacing traditional image processing methods which need to design features manually with end-to-end train detection and track segmentation algorithms on the basis of convolutional neural networks.And then,designing a joint detection network to simplify the structure and calculation.It achieved one shot detection from input images to output results,which is superior to the original network and traditional algorithms in accuracy and detection speed.The following is main content of this article:(1)Launching theoretical research on deep learning such as object detection and semantic segmentation.Combining with actual needs,we constructed an environment awareness scheme based on residual SSD and improved Seg Net network.The video from actual operation line is collected and the pictures are annotated.Train detection dataset and track segmentation dataset have been built to verify feasibility of the algorithms.(2)A train detection algorithm based on residual SSD network is proposed to realize train detection in various light environments.Considering that VGG-16,pre-structure of the original SSD network,has a weak ability to extract features from the input picture,the author chooses a 50-layer residual architecture as the basic network,then adds a transposed convolution structure to expand the receptive field of the feature map.In the prediction module,redesigned residual structure is established to further improve the detection accuracy.Through the actual test of trains in different environments and the comparison experiments of algorithm indicators,the effectiveness and robustness of the model in complex environments are verified.(3)Based on the encoder-decoder idea of the Seg Net network,the track semantic segmentation algorithm was redesigned.For the imprecise segmentation and hole phenomenon in Seg Net,a multi-scale cascade down sampling structure using multi-pooling technology was added to increase the weight of regional information.Thereby the feature extraction ability of down sampling is improved.Then introducing cascade dilated convolution to increase the receptive field.In the fitting optimization stage,the iterative endpoint fitting method is used to smooth the segmentation results,so as to output the flat track with a more realistic shape.(4)Based on the multi-task training idea,a joint detection model with shared feature extraction network is established.It combined object detection and semantic segmentation into a unified architecture,and experimental results are shown.The results show that this architecture has a satisfactory recognition effect on the train in forward track area,and meet the real-time and accuracy requirements of the train active environment perception.Figure 56,Table 11,References 74...