Visual-Inertial Odometry

This page presents research results from our visual-inertial odometry algorithms, implemented on a Samsung Galaxy SIII and an LG Nexus 4. The estimation is performed using only the phones' rolling-shutter cameras, 3-axis gyroscopes and the 3-axis accelerometers, and all processing takes place in real time on the phones' ARM processors. The estimates are computed without any prior maps, without using GPS, Wi-Fi or other external signals, and without any loop-closure detection or bundle-adjustment post-processing.

ICRA 2013 Demonstration
The first demonstration of our real-time system took place at the 2013 ICRA conference center, using a Samsung Galaxy SIII mobile phone. The video below shows the images recorded during a 1.1-km, 16-minute trajectory (the video is sped up 5x). The final position error is 5.3 meters, or 0.5% of the traveled distance, and the processing time on the device is 65 msec/image for feature extraction, and 11.4 msec/image for EKF updates.


Car Localization
The next video shows car localization in an urban environment, without the use of GPS. A Samsung Galaxy SIII is mounted on a car, during an 11-km, 21-minute long trajectory (video is sped up 10x, and also shows the ground truth collected by a separate GPS unit). Note that full 3D-pose estimation is carried out, but results are plotted in 2D. The estimation algorithm does not use any prior knowledge of the road layout or a motion model (e.g., non-holonomic constraints, almost planar trajectory). The largest position error is 0.57% of the traveled distance, while the processing time is 52 msec/image for feature extraction and 14.8 msec/image for EKF updates.


Mobile Robot Localization
The next video demonstrates 3D-pose estimation for a mobile robot, using an LG Nexus 4 device mounted on the robot. In this experiment, the robot moved on a 720-m, 11.7-minute long trajectory on the UCR campus. The final error is approximately 5.5 m, or 0.73% of the traveled distance, and the processing time is 46 msec/image for feature extraction, and 15.1 msec/image for EKF updates.