Research on Vehicle Adaptive Real-Time Positioning Based on Binocular Vision

Abstract

The positioning of a self-driving car is the basis of vehicle navigation and decision planning. To solve the problem of localization of self-driving cars in scenarios where GPS signals are missing, this article proposes an adaptive real-time positioning method, which is based on the binocular vision simultaneous localization and mapping method. The front-end visual odometry module uses the gray threshold adaptive oriented FAST and rotated BRIEF feature point extraction method for feature matching and pose calculation. The back end performs optimization processing based on the camera pose and the loop closure detection information estimated by the front end. Before optimization, we developed a rough dynamic feature elimination method to increase the robustness and positioning accuracy. Thus, a globally consistent positioning trajectory was obtained. To verify the actual positioning effect of the positioning system, the algorithm was tested using the KITTI data set and an experimental platform was built to test it in outdoor scenarios. The experimental results showed that the loop closure optimization and dynamic point elimination methods improved the positioning accuracy. The loop closure optimization significantly improved the actual positioning accuracy, and the dynamic feature elimination method slightly improved the positioning accuracy in dynamic scenes.