Vision-aided IMU for Handheld Pedestrian Navigation

Low cost inertial sensors are often promoted as the solution to indoor navigation. However, in reality, the quality of the measurements is poor, and as a result, the sensors can only be used to navigate for a few seconds at a time before the drift becomes too large to be useful. Therefore, it is necessary to regularly update the sensors with measurements from external systems such as GPS or other sensors useful for navigation. One such sensor is provided by the computer vision community where a camera can be used to obtain information about the relative translation and rotation between successive images. This paper describes the use of a camera attached to a low cost IMU for navigation in areas where GPS is unavailable such as indoors or deep urban canyons. It is assumed that a pedestrian user is walking with the mobile device held out in front of them with the camera pointing approximately towards the ground. Features are matched between successive frames, and the robust RANSAC framework is used to identify which of these lie on the ground plane, while estimating the camera's orientation and 3 dimensional body frame translation relative to its previous position. This information is used to aid the IMU using a Kalman filter to reduce the position drift. This paper describes the implementation of the combined computer vision and inertial navigation approach. A tactical grade IMU is used for initial testing since it provides more reliable measurements and enables us to provide a reference by which to compare the measurements obtained from the computer vision algorithm. It is demonstrated that even with a good quality IMU, the algorithm is able to significantly improve the performance of INS navigation when GPS measurements are unavailable.