Research On Key Algorithms Of Unmanned Vehicle Location Based On Visual Odometry

Unmanned Vehicle has become one of the hot topics in automotive research field,and how to realize the autonomous positioning for unmanned Vehicle is one of the key and difficult points.With the continuous development of computer vision technology,visual odometry has become an important choice for autonomous positioning.Only through the image information gathered by vehicle-mounted camera,visual odometry calculates the vehicle's 6 degree of freedom motion information based on the camera imaging model and the visual geometry model.This paper takes the basic principle of monocular vision odometry system as starting-point and focuses on the key algorithms of the odometer,including camera calibration,feature detection and matching,pose estimation.For the camera calibration algorithm,camera distortion and calibration error are not taken into account when calibrating camera based on the traditional two dimensional visual measurement algorithm.In combination with the basic principle of Zhang's plane calibration method,this paper presents an improved algorithm for camera interior parameter calibration based on two dimensional plane iterative optimization.The algorithm measures the internal parameters by setting the threshold,and the calibration results are optimized by the LM algorithm,which improves the calibration accuracy of the camera internal parameters.For the feature association algorithm,according to the requirement of real-time performance of unmanned Vehicle and the limitation of vehicle computing resources,this paper proposes an image feature extraction algorithm based on Harris and SIFT.This algorithm first detects the angular points of the image and divide the image into blocks according to the angular points distribution and extract features and describe using SIFT algorithm.The traditional SIFT algorithm and the algorithm in this paper are simulated on the Matlab platform.The performance of the two algorithms are compared in terms of the operating time of the algorithm,the characteristics of the scale transformation,the transformation of the visual angle and the sensitivity of the illumination.The experimental results show that algorithm calculation work can be considerably reduced by using the proposed algorithm without affecting the performance of the feature points.For the pose estimation algorithm,in order to solve the problem of poor real-time performance of the traditional 5 point pose estimation algorithm,a fast pose estimation algorithm is proposed.Based on the basic properties of the fundamental matrix,the fundamental matrix is solved by the epipolar geometry constraint.Then,the simulation experiment and the off-line experiment are conducted,followed by below steps.First of all,the accuracy and efficiency of the pose estimation algorithm is analyzed via simulations experiments.Secondly,the off-line experiment was carried out based on Chery's intelligent vehicle platform,and finally compare the motion estimation with positioning results of high precision GPS.The experimental results show that the complexity of the pose estimation algorithm is reduced without compromising the accuracy of the motion estimation results.