Research On Obstacle Intrusion Detection Method Based On Adaptive Change Detection And Self-supervised Mechanism

Obstacle intrusion in railway scene can endanger the lives of passengers and cause huge property losses.Timely detection of obstacle intrusion is conducive to ensuring the safety of railway train operation and passenger’s travel safety.The current obstacle intrusion detection methods have limitations in different aspects,so the obstacle intrusion detection in the actual field still relies heavily on manual inspection,which can not ensure the timeliness and reliability of detection.In this thesis,the change detection method is applied to obstacle intrusion detection,and a weakly supervised training method is studied under the limitation of labeling cost.To address the problems of camera shaking,vegetation shaking,environmental changing,and poor detection performance of small objects when applying the change detection method to obstacle intrusion detection task,a detection method based on adaptive change detection is proposed in this thesis.Firstly,an adaptive feature comparison module is designed to adaptively correct the comparison features according to the current camera shaking degree combined with the adjacent domain information,which reduces the false detection and missed detection of the camera shaking scene.Secondly,a difference feature correction module based on semantic relationship is designed to correct the outliers on the difference feature map by combining the semantic relationship of the reference feature map,which suppresses the interference of vegetation shaking on the detection results.Thirdly,the equal contribution difference fusion module is designed to improve the detection performance of the model on objects with different scales by reducing the information loss of difference features in the fusion process.Finally,an adaptive background updating method is designed to update the background in real time,which enhances the robustness of the model to the changes of light and shadow in the environment.To reduce the high labeling cost of dataset which required for model training of obstacle intrusion detection,low-cost scribble label is used to replace high-cost mask label,and a weakly supervised obstacle intrusion detection method combined with self-supervised mechanism is proposed in this thesis.Firstly,the initialization weight is obtained through pixel level contrast representation learning,which alleviates the over fitting problem of training the model with a small number of labels.Secondly,a self supervised consistency loss is designed to provide auxiliary supervision information for the training of unlabeled areas.Finally,a conditional random field loss based on difference feature is designed to make the model correct the detail information according to the long-distance relationship of difference feature map in the training process.Compared experiments were conducted on railway scene obstacle intrusion detection dataset,railway scene obstacle intrusion videos and four change detection datasets.The results show that the methods designed in this thesis can effectively reduce the missed detection rate and false detection rate,and are better than other obstacle intrusion detection methods and change detection methods.The effectiveness and generalization of the methods in this thesis for obstacle intrusion detection and change detection are verified.