Integrated estimation and control approach for vision aided inertial navigation

This paper discusses our research efforts in the area of integrated estimation and control of autonomous vehicles for vision-aided inertial navigation systems (INS) in the absence of globally defined markers (such as Global Positioning System (GPS) or known features). Camera-aided inertial navigation is effectively a bearings-only simultaneous localization and mapping (SLAM) application, were we are primarily concerned with accuracy of the unmanned ground vehicle (UGV) egostate (vehicle position, velocity and attitude) estimates and not the feature map itself. We simulate an UGV moving through a hallway, similar to other "urban canyon" environments, using a monocular camera to track unknown features to aid the INS estimation using a complementary form extended Kalman filter (EKF). It is well know that path choice plays a crucial role in estimator state observability, and thus in estimator accuracy. More accurate estimations lead to more efficient guidance, so it seems prudent to include estimation accuracy as a factor in guidance decisions. We demonstrate that for our implementation, covariance based N-step path planning would be unreliable. But, we find that simple path choices, such as sinusoidal or sawtooth, tend to do quite well without the express need for additional control and allow the UGV to accurately follow a desired path while maintaining strong egostate estimates for long stretches of hallway environments.

Integrated estimation and control approach for vision aided inertial navigation Conference Paper