Research On Real-time Perception Technologies Of Track Environment For Unmanned Aerial Vehicle

With the increasing mileage of railway transport in China,especially in the complex geomorphological environment in the western areas,the reliable guarantee of rail transport operation security becomes more and more difficult.The railway intrusion seriously affects the safety of train operations.Video Surveillance is currently the main method for track environment perception.Due to the inherent limitations of surveillance cameras,it can only be deployed at a fixed position along the track lines to prevent and control key areas.In order to relieve the increasing pressure of rail traffic operational security and ensure the rail transit safety guarantee technology to meet the requirements of the rail transit industry,the China Ministry of Science and Technology plans to take the real-time identification of rail transit operating environment based on mobile platforms as the core research content in the "13th Five-Year Plan" national key R&D project "The Comprehensive Guarantee of Rail Transit Operation and safety based on Space,Sky,Train and Ground Information Integration" and "railway system security assurance technology".In recent years,the UAV-based target recognition has made great progress in industry,agriculture and national defense construction.However,due to the extremely complex surrounding environment,the UAV-based rail transit safety detection method is still checked by human source review,and has not been achieved automatic recognition so far.This paper focuses on the key technologies of UAV-based track environment real-time perception systems.The details are as follows:Aiming at the jitter problem when UAV collects videos,a video stabilization algorithm based on improved Harris corners and KLT feature point tracking is proposed to obtain a stable and clear track environment video sequence.Aiming at the problem of track region of interest(ROI)extraction,an improved Canny edge detection method is proposed,which greatly eliminates the influence of surrounding environment when extract current track edge.Probability Hough change and Kalman filter tracking method are also be used to accurately locate the track position.According to the rules of track clearance,the region of interest in the current image is successfully extracted.For the intrusion detection problem in the track ambient environment,this paper proposes an intrusion detection method based on the fusion of geographical location information and track visual feature database.In the construction phase of the feature database,the geographic coordinates are used as indexes and advanced ALP feature descriptors are used as the track visual features.The track ALP visual features have the characteristics of stabilization,strong robustness to environmental illumination,angle changes,and et.al.Finally,this paper proposes an intrusion detection method based on visual saliency detection.Intrusion regions are accurately extracted by integrating track gradient saliency maps,Lab color space saliency maps and morphological methods.Experiments show that the UAV-based track environment perception technology can achieve a detection success rate of 96%for intrusion detection;the range of small object intrusion and multi-target intrusion on the track can also be detected and identified effectively.Each frame processing speed of the recognition satisfies the inspection speed requirement 40 km/h of the rotorcraft UAV.