Research On Image Semantic Segmentation Method For Complex Environment Of Railway Shunting Operation

With the rapid development of China’s railway system and the continuous expansion of the railway network,shunting operations,as an important part of railway operation,are facing tremendous challenges.In the process of shunting operations,railway workers need to operate in complex environments such as stations and freight yards.Therefore,achieving automated shunting in such complex environments is a challenging problem.In recent years,various algorithms based on deep learning theory have been widely used in fields such as face recognition,object detection,and autonomous driving.However,their application in railway safety is relatively limited.This paper applies the image semantic segmentation method based on deep neural networks to the field of railway shunting operations.We conduct research on the detection of locomotives,pedestrians,vehicles,railway switches’ running status,and running direction based on deep neural networks.This helps provide intelligent analysis and decision-making support for railway operation safety and provides reliable guarantees to eliminate hidden safety hazards in railway operations.The main work of this paper is as follows:Firstly,starting from modeling context information and utilizing human visual-inspired attention mechanism,we propose an attention-guided real-time semantic segmentation network AR-SS for complex environments in railway shunting operations.Firstly,we use Mobile Netv2 as the backbone feature extraction network to speed up the model processing speed and reduce memory consumption.Secondly,we use self-attention mechanism to improve the correlation between pixels at long distances and promote the fusion of global features.Thirdly,we use cross-layer feature fusion to integrate the detail information of multiple feature maps and improve the model’s ability to express detail information.Finally,we introduce a global feature upsampling module to better combine high-level and low-level semantic information and enhance the image’s detail expression ability.Experimental results show that our model achieves a segmentation accuracy of 82.1% and a forward inference speed of 59 FPS.This method achieves a balance between segmentation accuracy and inference speed,and it is an efficient approach for solving complex environment tasks in railway shunting operations.Secondly,from the perspective of lightweight model design,a lightweight real-time semantic segmentation model called Ms M-SS is proposed based on the orbit image semantic segmentation algorithm using self-attention mechanism described in Chapter 3.Firstly,Mobile Net is employed as the backbone network to reduce model prediction time and memory overhead.Secondly,stripe pooling is introduced to construct a hybrid pooling structure instead of the original dilated spatial pyramid pooling structure,establishing long-range dependencies for global information.Finally,the decoding layer of the model from Chapter 3 is improved by incorporating multi-scale feature information,enhancing the correlation of information across different levels and improving the expressiveness of image details.Experimental results demonstrate that compared to the AR-SS model,this approach further reduces the number of model parameters,achieves a forward inference speed of 67 FPS,and a segmentation accuracy of 82.8%.This article deeply explores the semantic segmentation of images in the complex environment of railway shunting operations,from the perspectives of lightweight model design,context information modeling,and attention mechanism.It enriches the research content in the field of railway operation and shunting safety and provides technical support for the automation of railway shunting operations.