Research On Obstacle Detection Algorithm In Front Of Train Based On Image And Radar Data Fusion

Ensuring the safety of train operation is the prerequisite for the normal operation of railway.The collision between the train and the object ahead is the main type of railway safety accidents.For reasons of safety of operation,it is required to apperceive the running environment.The existing and in-use detection methods mainly include manual line patrol,video monitoring at accident prone points,or erecting protection network at mountain bridges and tunnels.Obviously,they can not meet the demand.In this thesis,the image and radar data fusion method is proposed to detect the running environment in front of the train,which can detect the running environment in real time,find the obstacles in front timely and effectively,and reduce or avoid the consequences of collision accidents.It can make up for the disadvantages of limited visual distance,small fixed-point monitoring range and insufficient reliability of single sensor detection,and has great practical value to ensure the safety of train operation.Through the detailed analysis of the railway operation environment,the algorithm proposed in this paper is divided into three modules.The main contents of this paper are as follows:This thesis analyzes the video image in front of the train taken by the vehicle camera,and designs a rail detection method based on LSD(Line Segment Detector)and the least square curve fitting.After that,a smooth rail curve model in pixel coordinates is obtained by using the idea of multi-line approximation curve.Based on SSD(Single Shot Multi Box Detector)algorithm,a deep learning network for object recognition is built.The set object is fully trained,and the captured image is substituted into the network for object location and recognition.It is concluded that there are two kinds of intruders,one is unilateral intruder and the other is penetrating intruder.The corner coordinates of the object recognition positioning frame and the image boundary curve are compared to get the judgment result of the object intruder.The radar error correction scheme is designed by measuring the target at the calibration point several times and the measurement characteristics of the FD4-400CJ10 radar used in this research are analyzed.After error analysis,the measured longitudinal distance with smaller error is not corrected.Taking the measured longitudinal value as reference,the transverse measurement error has a strict linear relationship with the longitudinal distance within the range of ±7.2° of horizontal angle.Combined with the limit standard and error compensation,the intrusion detection area in radar coordinate system is constructed,and the target in the area is identified as the intrusion object.Based on the GPS on-aboard,the plane coordinates of the measured GPS data are obtained through the transformation from WGS-84 coordinates to Gaussian plane coordinates.Through analyzing the causes of GPS error data,the outlier elimination scheme is designed,and more accurate data belonging to the line is obtained.The complete line is divided into curve line element,straight line element and line to curve transition line element.The GPS measured transformation coordinates are divided into different line segments after slope difference classification,and the segmented line element model is obtained by the least square.The boundary curves of the segmented line element are calculated by the lateral error of radar detection and the gauge standard.The radar locates the line map through GPS data,reads the clearance data in front of the line point,and compares the target data with the clearance through coordinate transformation to obtain the real-time detection results of the object.By comparing the time efficiency of machine vision and radar,taking the radar data frame as the reference,the latest time matching is used to match the image detection results with the radar detection results.Through the analysis of imaging principle,camera calibration and position measurement,the mapping model from radar to matched image frame is constructed.A 17:18 ROI of radar mapping point is set,and the overlap area ratio between the visual detection result and the ROI is calculated to judge the consistency of the radar detection target and the image detection target.The fusion detection results are divided into two risk levels.The test results show that the multi-sensor fusion detection is more applicable than single sensor detection,which overcomes the shortcomings of machine vision failure in light shortage and makes up for the defects of single radar detection that can not recognize the object types.In this test scenario,the comprehensive detection rate is about 88.67%,the missed detection rate is about 4.28%,and the detection performance is better than that of single sensor.