Research Of Obstacle Detection And Recognition Technologies In Cross-Country Environment For Unmanned Ground Vehicle

Unmanned ground vehicle has a potential applications in many fields.As a main contentof UGV research,the obstacle detection and recognition technology has been a hot research,but also difficulty,especially for obstacle detection and recognition in cross-country environment.Thus,an obstacle detection and recognition method based on multi-sensor data fusion is proposed in this paper to realize the cross-country environment.The specific contents are as follows:Firstly,aiming at the system of camera and LIDAR we use,a new joint calibration method based on plane feature is studied and proposed in this paper,and this method can also be used in LIDAR calibration.This paper completes the camera calibration by Zhang's method,and with the new method,the LIDAR calibration and the joint calibration are completed.Judging from the verification of test,the accuracy of this calibration method can reach 1.2931 mm,and it provides reliable basis for confirming the location of obstacle.Secondly,In the process of extraction of ROI(region of interest),for the extraction of ROI based on visual image,a method of color image segmentation based on HSV and Lab color space is proposed in this paper,and this method uses the fisher criterion function segmentation in HSV color space and K-means clustering segmentation in Lab color space.The experiment result shows that this method can locate precisely the obstacle region and assures the versatility and real-time of image segmentation.For the extraction of ROI based on laser data,this paper transforms the laser data into a height-gray image and gets the segmented image by using K-means clustering segmentation in height-gray image.After segmentation,the final ROI based on visual image and laser data can be extracted by the following morphological treatment.Thirdly,this paper explores the discrimination of different features for obstacles,and four different features with high recognition are extracted,including S/V feature based on HSV space,b-a feature based on Lab space,?3/?2 feature based on covariance matrix of laser data,penetrability value feature based on laser data.In the meantime,each feature's eigenvalue interval of different obstacles was calculated by analyzing a great number of training samples.The feature extraction has lay a good foundation for obstacle recognition.At last,a new obstacle recognition method based on D-S evidence theory was proposed,and it is verified by the simulation model put up using MATLAB.Using this simulation model,the recognition results based on single feature and D-S evidence theory can be obtained respectively.According to recognition results of lots of samples,the accuracy of recognition method based on D-S evidence theory reaches 88% and it is much higher than the method based on single feature.At the same time,this paper discusses the localization of obstacle based on joint calibration,and the result of discussion shows the necessity of joint calibration.All simulation results indicate that All the simulation results show that the proposed method has good accuracy,real-time and robustness.