Research On Domain Adaptive Object Detection For Visual Monitoring Of High-speed Railway

Intelligent monitoring based on computer vision is widely used in the field of rail transit because of its high-precision,automation and non-contact characteristics,especially the high-speed railway visual monitoring application with fast speed and less skylight time,which requires all-weather real-time monitoring of key facilities,and detect and locate key targets from the monitoring images and perform intelligent analysis and processing to provide guarantee for the safe operation of railway.However,with the increasing number of high-speed railway lines and the upgrading of monitoring equipment,it is difficult for the detection model of a certain route to take into account the detection tasks of multiple lines,and relabeling new data is time-consuming and laborintensive,Therefore,the research of detection models with stronger adaptability and mobility has important significance and practical value for improving the efficiency of high-speed railway visual intelligent monitoring.For the above problems,based on the domain adaptation technology,this paper studies how to use the labeled line monitoring image data(source domain)and unlabeled line monitoring image data(target domain)to improve the object detection model on the new line(target domain).This paper considers the characteristics of high-speed railway visual monitoring images,and puts forward a corresponding domain-adapted target detection model based on the focus on the shallow feature instance alignment method and the multi-scale feature fusion strategy,and conducted experimental analysis and discussion on the public data set and the high-speed railway displacement monitoring data set of our project.The main research contents of this paper are as follows:(1)A domain adaptation detection model combined with shallow feature instance alignment module(Shallow Alignment Module,SAM)is proposed.Current domain adaptive detection methods based on adversarial ideas usually align the global feature level and the individual instance feature level to extract common features in the two domains.However,for the individual instance level,shallow features are underutilized,especially for image samples with rich shallow features(such as high-speed rail displacement monitoring images with obvious edges and texture features).Therefore,this paper proposes an instance alignment module based on shallow features,adding a domain classifier to shallow instance features,and using gradient reverse layer(GRL)to counter training to further optimize domain invariant features.The experimental results on the public datasets of Cityscape and Sim10 k show that compared with the SWDA model,the improved model achieves 2.0% and 1.4% m AP improvement respectively,and compared with the HTCN model,the improved model achieves 1.2% and 0.8% m AP improvement respectively.(2)A domain adaptive detection model based on Multi-Scale Feature Fusion Network(MFFN)is proposed.Taking into account the multi-scale characteristics of targets in different domains,especially the large changes in target scales in monitoring images of different lines,inspired by the feature pyramid network(FPN)fusion method,a multi-scale feature fusion network oriented to domain adaptive detection is proposed.Utilizing the high-resolution features of shallow features and the rich semantics of deep features,fusion of multi-layer and multi-scale features are input to the region proposal network(RPN)to improve the accuracy of domain adaptative detection.The experimental results on the public datasets of Cityscape and Sim10 k show that compared with the SWDA model,the improved model achieves 0.9% and 1.5% m AP improvement respectively,and compared with the HTCN model,the improved model achieves 1.1%and 0.8% m AP improvement respectively.(3)On the basis of the above two researches,a domain adaptive detection model for high-speed rail visual monitoring is proposed.The model combines the shallow feature instance alignment module,makes full use of the texture and edge information of the monitoring image,uses a multi-scale feature fusion network to provide a more informative fusion feature map,and improves the anchor frame setting scheme according to the morphological characteristics of the monitoring object.The experimental results on the high-speed railway visual monitoring data set show that compared with the SWDA model,the model proposed in this paper achieves a 1.4% AP improvement on the highspeed rail visual monitoring data set.