Research On Robust Detection And Tracking Method Of Railway Foreign Object Intrusion Based On Machine Vision

The cross-development of China's railways has brought new challenges to the construction of railway traffic safety guarantee system.Due to the non-closed design of railway lines and supporting facilities,airborne intrusion such as kites,plastics,tile and kongming lanterns are easily blown into the railway limit by the wind,if wrapped in the key parts of catenary like messenger wire,hanging string and contact wire,train set need to pass slowly or stop,in severe cases,catenary and pantograph may be damaged,caused many trains to be late.On the other hand,when the train is running at a high speed,it will generate a high negative pressure,ground intrusion like pedestrians,animals and vehicles will be drawn into the driving limits,bring a great threat to the rail traffic safety.How to detect and track rail intrusion of foreign object is a hot issue in the field of rail power supply and transportation.This article is based on the National Natural Science Foundation of China and Cultivation Program of Science and Technology Department of Jiangxi Province,based on the existing railway integrated video surveillance system,a scale-adaptive high-speed foreign object intrusion detection and tracking method based on machine vision is studied,provide technical support for the development of the intelligent rail transit industry.The main research contents in the detection of foreign object intrusion are: a visual background extractor is used to create a background texture model for the video sequence,and the pixels based on the background color model and the background texture model are all determined as foregrounds,which can effectively reduce the false detection rate of camouflage color intrusion.In the tracking phase of foreign object intrusion,this paper has innovated from the following three aspects.1)Obtain block cycle position training sample matrix by dense cyclic sampling of target region,project a position translation training sample set to a linear separability high dimensional space,according to the diagonalization of the block cycle matrix in the Fourier domain,the inversion operate in the linear regression normalization equation is simplified to point multiplication,significantly improved the tracking speed of algorithm.2)Extract Histogram of Oriented Gradient feature and Color Name feature of foreign object,training two kernel correlation translation filters respectively,perform linear weighted and filter adaptive update according to the Smooth Constraint of Confidence Maps and average peak-to correlation energy,it effectively solves the problem that the tracking frame is easy to drift under the scene of target deformation and illumination changes.3)It is easy for the algorithm to learn excessive background or local texture information when tracking multi-scaled object,which causes filter pollution,training a scale filter with the FHOG feature of initial tracking box,construct a scale test sample set as a scale filter input using the scale pyramid technique,and the test sample with the highest output response is used as the scale of the foreign object of the current frame,the algorithm can be used for continuous tracking at critical ultra-close ranges.Finally,a series of railway foreign body intrusion video sequences collected by the railway integrated video surveillance system along the railway are used as experimental examples to verify the proposed algorithm,The experimental result shows that the proposed PSA-Kcf algorithm outperforms the Mean Shift algorithm and the native kernel correlation filter algorithm without scale estimation in tracking accuracy,slightly higher than SA-Kcf and SAMF algorithm;the tracking speed is faster than Mean Shift,SA-Kcf and SAMF algorithm.Equivalent to the Kcf algorithm without scale estimation.