Research On IMU-aided Stereo Visual Odometry

Visual odometry is a technique for robots to mapping surrounding environment and determine their positions and orientations while moving.Visual odometry is an important component of visual simultaneous location and mapping(V-SLAM)and can be interpreted as a V-SLAM without loop closure.With the promotion and popularization of mobile robots,visual odometry is becoming a popular research in computer vision and robotics.Inertial measurement unit(IMU)is a device can measure its linear acceleration and angular velocity,and usually combines other devices to measure position and orientation with high frequency.In this thesis,we proposed a visual odometry implementing direct methods,which uses images from stereo camera and combines IMU to estimate images,positions and orientations.Direct methods solve relative positions and orientations through minimizing the photometric error between images.Compared to feature-based methods,direct methods don't calculate feature descriptors and can use more information in images.This makes direct methods more accurate and robust.The main research includes:1.Use photometric transfer parameters to calculate photometric errors.And add them to the states of images,optimize them with images' positions and orientations.These make photometric error calculation more robust.2.The errors calculated by stereo image pairs and the errors calculated by temporal image pairs are optimized together.This makes images' positions and orientations constrained by different depth range landmarks.3.The errors related to IMU and the errors related to images are tightly coupled and optimized together.Also,we constrained the optimization in a small window to ensure instantaneity.