Research On Key Techniques Of High-precision And Real-time Visual Localization

Localization algorithm based on machine vision is a hot topic in the field of intelligent mobile robot. By analyzing consistent information of the static scene included in the image sequence, visual localization can help the mobile robot obtain the current position and orientation information accurately, ensuring that the mobile robot can accomplish the appointed task safely and efficiently.This dissertation aims at high-precision and real-time visual localization of the mobile robot under large-scale complex environment, and presents a critical study centering on three key issues of accuracy, real-time and robustness. In order to precisely estimate the robot’s six degrees of freedom (DOF) measurements in position and orientation, a binocular stereo visual odometry (VO) is built based on the matching of feature points. To improve the positioning accuracy, a Two-stage local binocular bundle adjustment (TLBBA) is proposed in Chapter2, which optimizes the motion estimation results by taking full advantage of the information and constrain in the binocular image sequence. Chapter3implements a real-time binocular VO, through designing the algorithms and modules of the system reasonably and making best of the computing resource and parallelism of algorithms. In order to increase the robustness of local motion estimation under the complex environment, Chapter4proposes a compressive feature based on an adaptive multi-feature appearance model, by which the image patches along the robot’s heading direction can be tracked and the yaw angle can be estimated robustly to correct the unreliable motion estimation results. To suppress the accumulated localization error when the mobile robot travels in large-scale environment, a global position correction algorithm is proposed based on building and matching panoramic compressive landmarks online, and the path drifting problem can be solved efficiently. The main content and innovation of this paper are as follows:1. A two-stage local binocular bundle adjustment optimization algorithm is proposed, which makes the best of the consistent information in image sequences and improves the precision of the binocular VO based on feature point matching. In the first stage, the single step motion estimation result is optimized. Compared to the maximum likelihood estimation based on the uncertainty of three-dimensional points, this algorithm uses the uncertainty of two-dimensional feature points whose error distribution is more uniform. Moreover, the binocular model is introduced for the more plentiful and reasonable constrains. In the second stage, motion parameters and structures in a sliding window are optimized simultaneously, which also benefits from the binocular model and uncertainty of two-dimensional feature points. Compared to the traditional monocular LB A, our TLBB A has more accurate initial value, more reasonable object function and higher precision.2. A real-time binocular VO system is presented. In order to improve the processing speed without losing accuracy, our VO system extracts SIFT features based on GPGPU, controls the matched features based on a grid matching method, and estimates the motion parameters using RanSaC and the least square method of HORN with GPGPU acceleration. At last, the whole system is divided into two parallel threads:the feature matching thread is responsible for feature extraction, matching and three-dimensional rebuilding. The motion estimation thread is responsible for motion estimation, TLBBA optimization and computing the robot’s global position by accumulating the single step motion parameters.3. An algorithm for local yaw angle calculation is proposed based on the adaptive multi-feature compressive tracking of image patches. The change of the mobile robot’s yaw angle can be calculated more robustly with this algorithm, when the binocular image sequences lack consistent features. To improve the precision and real-time of the image patches tracking, an adaptive multi-feature appearance model in compressive space is proposed:a sparse two-stage random measurement matrix is constructed to compress the SURF feature, making the original intensity-based visual representation have higher accuracy and stronger descriptive ability. The weighting factors of features in the statistical model are adjusted adaptively by analyzing the features’ ability of discriminating the objects from background. So the redundant and useless features are suppressed and the efficiency and accuracy of statistical model are improved.4. A global position correction algorithm is proposed based on the online panoramic landmarks online. The panoramic images are described with the adaptive compressive features of two scales, and the best matching of the current image can be found in the landmark library quickly and accurately. With the detection of road crosses, the utilization and matching accuracy of visual landmarks are improved, and the matching cost is further reduced. When a landmark is revisited, the robot’s position can be corrected after the motion estimation between the current image and landmark image. Then the accumulated localization error can be reduced greatly, and the precision and robustness of the localization system under large-scale complex environment can be improved.