Research On Pedestrian Detection Technology Based On Fusion Of Radar And Machine Vision

With the innovation of sensor and computer hardware manufacturing,the continuous development of image processing and artificial intelligence technology,pedestrian detection has become a rapidly developing and important field in machine vision.In pedestrian detection,the surrounding environment information is acquired by the sensors to search for the presence of pedestrians in the detection area.Currently,pedestrian detection is mainly used in smart vehicles,unmanned driving,video surveillance and robot interaction technologies.In traditional detection methods optical sensors based on camera are mostly used.Optical sensors are susceptible to light and weather conditions,and in order to obtain a higher detection rate,time is sacrificed to process complex image data.Unlike the flaws of camera,the radar sensor has the better detection reliability and robust.However,due to the simple type of radar data,the radar is not enough for pedestrian recognition.Therefore,the research of this thesis is to fuse radar and camera to make up for their shortcoming to realize pedestrian detection,and finally detection accuracy and efficiency is improved.Firstly,in the radar module the Walabot radar is programmed to trigger and collect the three-dimensional data of the detected target,and then the empty and invalid data is removed.Aiming at the noise brought by the measurement process,Kalman filter is designed to apply to noise filtering.And combined with the prediction range in prediction process of Kalman filter,some radar targets with larger errors are removed.After experimental analysis,filtering can reduce noise interference and improve the stability of radar detection.Secondly,in the camera module the three-frame difference method is used to extract the foreground information of the moving target,and the mathematical morphology method is used to filter the foreground image noise,and then the target contour and the candidate window are obtained.Combined with the approximate feature pyramid multichannel pedestrian features including color,magnitude and direction of gradient can be obtained to train the XGBoost classifier in multiple iterations.Then this thesis uses the soft cascade method to optimize the XGBoost classifier to realize the rapid recognition of non-pedestrians and further improve the detection efficiency.Finally,in the fusion processing module a fusion detection scheme of radar and camera is designed.After completing calibration and synchronization of sensors in space and time,the pedestrian similarity of the candidate window formed by the frame difference method will be analyzed to filter the window.Then these candidate areas containing moving pedestrians are fused with the candidate window generated by radar based on the target distance to obtain the final window for classification and recognition.At the same time,this thesis uses the Soft-NMS method combined with the target distance to improve the screening of the result window.Based on the Open CV open source library and C++ programming environment,a hardware platform and software environment are built to verify the pedestrian detection scheme.The results show that the single-frame image detection time is reduced after fusion and the precision rate is also improved.