Stereo Visual Odometry With Automatic Brightness Adjustment and Feature Tracking Prediction

Abstract

Vision-based localization and mapping can be easily affected by unstable feature tracking and illumination variations. To address these problems, we propose a point-based stereo visual odometry (VO) system with image brightness adjustment and feature tracking prediction. The system incorporates two threads that run in parallel: front-end and back-end. The front-end thread performs brightness adjustment, feature tracking, and motion estimation between frames. When the brightness of image changes significantly, a cumulative gray-scale histogram is used to estimate the exposure of the camera and adjust the brightness of the image. Additionally, a constant acceleration motion model and stereo geometric constraint are used to predict the location of feature points in the target image, providing a reliable initial guess for the Lucas–Kanade (LK) optical flow tracker. In order to improve the accuracy and reduce computational complexity, the back-end performs a sliding window bundle adjustment (BA) to achieve optimal camera poses and landmark positions. Experiments on publicly available datasets indicate that the proposed scheme has a better performance than state-of-the-art stereo VO.