Robust state estimation for real-time quad-rotor using loosely coupled VINS
In this thesis, our goal is to figure out a system solution of robust state estimation that can run on quad-rotor platform in real-time. The basic module is stereo visual odometry (SVO), a relative observation for determining the pose change along with time. Two fisheye cameras are used and a camera model is built to calibrate the images so that the stereo matching and tracking frame to frame can be done on rectify images. Improvement of calibration, detection and tracking process are applied and performance of accuracy is analyzed with both benchmark dataset online like New Tsukuba Stereo Dataset and real time dataset collected with the ground truth provided like Vicon motion capture system. To make the result robust and accurate, Inertial Measurement Unit (IMU) data is also incorporated and fused with SVO output using Unscented Kalman Filter (UKF). So together the method we use is the LCV, loosely coupled VINS (Vision-Inertial). We are using Robot Operating System (ROS) as the interface and the software is made in module and can be extended to take in information from other sensors like GPS and Laser scan. The robustness is illustrated and quantized in a variety of indoor and outdoor environments .
