Keyframe-based stereo visual-inertial SLAM using nonlinear optimization

Accuracy is highly important on autonomous robots. In this work, we propose a novel visual-inertial SLAM with stereo camera and IMU, which construct sparse map and estimate the camera poses accurately. The camera and IMU data are tightly coupled by nonlinear optimization. pre-integration is used to integrate rotation, velocity, and the pose matrix. A serious techniques are adapted to feature extraction, keyframe selection select keyframes, and loop closure. In addition, the system can run real-time on standard computer. The system localization accuracy can arrive centimetre-level especially in a large scale environment, and system is robust. We elevate the system on public datasets to compare other visual-inertial SLAM approaches; our system achieves better accuracy and robustness.