Real-Time Pose Estimation by Fusing Visual and Inertial Sensors for Autonomous Driving

Abstract

The thesis works on developing a visual inertial odometry with feature-based stereo visual frontend, IMU preintegration and sliding window nonlinear optimization backend. The fusion of vision sensors and inertial sensors provide a robust and complementary pose estimation system to overcome the weaknesses of the vision-only or IMU-only systems. Stereo camera is easier to acquire the depth compared to the single view in monocular one, meanwhile has larger range and cheaper price than the RGB–D camera. When considering data association between several images, the ORB methods are applied to trade off between computational accuracy and efficiency. The inertial measurement preintegration is implemented to transform a number of measurements between selected keyframes into single relative motion constraints, which provides efficient handling of high rate sensor inputs. The problem as a whole is constructed as a bundle adjustment problem with motion constraints. It is treated as least square problems and solved with the LM method. The system performance is evaluated through EuRoC datasets, considering slow and fast motion, bright and dark scene. The average processing speed falls in the range of 19 to 26 FPS. The minimum translation and rotation error can respectively reach 0.11 m and 1.39 . Overall this work proposes a new VIO framework, examines its performance by evaluating open datasets as well as comparing to the existing VIO algorithms, and also brings up discussions on possible directions for further improvement.