Research On Robust Instance Segmentation Algorithm In Rail Transit Scenarios

The objective of instance segmentation is to detect the independent instance from the image and mark the pixel region of the object,which is one of the basic tasks in the field of computer vision.Related research has important theoretical significance and wide application prospect.With the rapid development of rail transit,ensuring the safe operation with computer vision technology has become an important demand.Therefore,this paper carried out relevant research on instance segmentation for rail transit scenarios,combined with attention mechanism and contrastive methods to improve instance segmentation performance and built a rail transit instance segmentation dataset.The main works of this paper are summarized as follows:(1)An instance segmentation model with spatial relationship constraints is proposed.Firstly,the image features are extracted by the backbone network enhanced by spatial information,and then the features are fused by a multi-layer feature pyramid network.Then the feature is divided into multiple equal size regions,and the spatial information is learned through the constructed spatial relationship constraint network to enhance the feature representation.Finally,the proposed model uses multi-scale region features to generate instance segmentation results.On COCO dataset,compared with SOLO model,the m AP index of the proposed model improved by 0.3%.Experimental results verified that spatial relationship can improve the instance segmentation performance.(2)An instance segmentation model based on contrast enhancement is proposed.In this model,a memory network and a prototype network are designed to improve the performance of instance segmentation.The memory network can make full use of interrelationship between features to enhance the ability of feature representation by enlarging number of features available for reference,storing features and constructing positive and negative pairs.The prototype network uses feature to learn category related semantic prototype,and then segmentation model training with image features to improve instance segmentation performance.In COCO dataset,compared with YOLACT model,the m AP index of the proposed model improved by 1.3%.Experimental results show that the contrast enhancement method can make full use of intra-sample relationships to improve the instance segmentation performance of the model.(3)An instance segmentation model for rail traffic scenarios is proposed.Aiming at the risk warning requirement,an instance segmentation dataset of rail transit scenarios is constructed.Combined with the above spatial constraints and contrast learning networks,an instance segmentation model for rail transit scenarios is constructed.Based on the segmentation results,the proposed model combined with the prototype network to identify the anomaly instance segmentation under the open set setting.In the CAOS anomaly segmentation dataset,compared with Synth CP,the m Io U of the proposed model is improved by 0.3%.The accuracy of the constructed dataset of rail transit scenarios reaches 96%,and the experimental results show that the proposed model can be applied to the risk warning task of rail transit scenarios.