| In order to improve people’s happiness index and satisfy people’s yearning for a better life,in the municipal transportation options,major cities focus on subway trains in urban rail transit.As a hot spot of people’s attention,driverless technology has been applied to devices such as driverless cars and drones.Subway trains have learned from this technology to realize driverless trains,and laid nine driverless lines in China.Driverless subway trains are the trend.However,obstacles often threaten the normal operation of trains.In response to this problem,CRRC took the lead in launching the development and test project of standardized Chinese subway trains.After investigation,discussion and summary,the two parts of the subway train obstacle system,the active obstacle detection system and the passive obstacle detection system,were respectively proposed.specific technical requirements.This thesis studies the current active obstacle detection methods in unmanned driving technology at home and abroad,referring to the characteristics of subway trains in the process of operation,in order to meet the technical requirements of active obstacle detection in subway train obstacle detection systems(Detecting the train in front within 300 meters,pedestrians within 150 meters,toolbox within 50 meters),put forward the construction plan of the active obstacle detection system platform: In order to get rid of the limitation of detection distance,the system uses color cameras with long and short focus lenses as vision The sensor collects road condition information.Based on neural network learning,the system can provide early warning to obstacles in the warning area ahead of the train.The system uses the YOLO algorithm idea in the neural network and the tower-shaped enhancement feature extraction scheme to process the collected image information.After training,select the optimal model MAP value of 90% to accurately identify the obstacle category,location,credibility information in the image.Then,the system draws on the Deep Lab semantic segmentation idea and cooperates with the spatial hole pooling idea.The m Io U and m PA values of the optimal model after training are both above 94%,and the alert area in the image information can be accurately divided.The system integrates the two information,and only provides early warning for obstacles located inside the warning area.After theoretical analysis and research,according to the braking principle of subway,an active obstacle detection method based on neural network learning is proposed.The entire obstacle detection system includes the system host,long and short focus cameras,and the main box equipped with the system.According to the scheme of the active obstacle detection system proposed in this thesis,an active obstacle detection platform for subway trains is built.The hardware required for the platform includes vision sensors(color cameras and long and short focal lenses to overcome distance limitations)and system hosts.After comparative analysis,Hikvision Robot’s MV-CA060 series cameras were finally selected with 16 mm and 50 mm focal length lenses.The graphics processor of the system host adopts the RTX-A4000 series graphics processing unit of NVIDIA.Finally,based on the established subway train active obstacle detection platform,this thesis uses the active obstacle detection technical indicators as the test standard to verify the actual scene of the obstacles,and the results prove the feasibility of the platform built in this thesis. |
