Key Technology Research On Automotive Environment Monitoring System Based On Panoramic Vision

Panoramic vision based automotive environment monitoring system can provide two driving assistance functions for the driver,as panoramic imaging and objective detection.It is an important technology in the field of active safety and has important application value.Existing target detection is based on normal camera or radar,and its detection algorithm does not apply to fisheye camera with large visual field and large distortion.This paper focused on research of pedestrian detection algorithm,based on panoramic vision,and improvement of the panoramic imaging system calibration methods,in order to obtain more accurate system parameters.Calibration principle of panoramic imaging system was analyzed.By comparing experiment,Catadioptric model was chosen as the imaging model of fisheye camera,based on which the calibration of cameras and the panoramic imaging system was conducted.A stereo calibration board was designed for fisheye camera,consisting of three mutually perpendicular boards,so that the corners could completely cover the camera to make full use of the fisheye image’s edge portion,in order to obtain more accurate camera parameters.A positioning block was added to the calibration board to ensure the stability of the system calibration.Front projection and histogram equalization was conducted on the fisheye image as pre-processing,for image feature standardization.Normalized pedestrian sample set was built.Normal pedestrian feature descriptors and machine learning algorithms were analyzed.Classifier training experiment was conducted using AdaBoost method combined with Haar features,and SVM method combined with HOG features,as well as Convolution Neural Network.Classifiers and the training process were evaluated to summarize their characteristics.Partial HOG features and entire HOG features of pedestrian were compared,and SVM classifiers were trained using ROI-HOG feature.A CNN classifier was designed after optimize the parameters.Combining unsupervised CNN feature extraction and linear SVM as supervision,a CNN-SVM pedestrian classifier finally came out,which could realize fast and accurate pedestrian detection.With SVM classifiers based on ROI-HOG features as weak classifiers,strong classifiers were built by AdaBoost method,to form the final cascade classifier,which could achieve fast and accurate pedestrian detection.To make full use of the large visual field of fisheye camera,huge distortion on the edge of the image had to be taken out.Front view projection method with multi yaw angles was proposed.One fisheye image was transformed to several front view projection images with different yaw angles for pedestrian detection.Pedestrian in any positon of the fisheye image could be reshaped back to normal with affine deformation removed,in the front view projection image with some yaw angle,to facilitate the subsequent detection.Setting rules of the yaw angle were summarized after experiment to minimize the number of projection images needed,to reduce detection time consuming.Panoramic calibration experiment was conducted on a car.Experiment showed a good performance of calibration accuracy and stability.Videos were captured to build the test dataset,by marking pedestrian position frame by frame.Evaluation software was designed on PC.Pedestrian detection program was run on it to evaluate the algorithm’s performance.Experiment showed that a high pedestrian detection rate was achieved within a certain range of distance.