Feature and pose constrained visual Aided Inertial Navigation for computationally constrained aerial vehicles

Detecting and dealing with hovering maneuvers in vision-aided inertial navigation systems

“…Note also that, without loss of generality, we express the image measurement in normalized pixel coordinates, and consider the camera frame to be coincident with the IMU frame 3 . By differentiating the nonlinear measurement model (14) with respect to the augmented state (9), we obtain the linearized measurement Jacobian:…”

“…Recent work on VINS, addresses the case of hovering by utilizing hybrid filter estimators that include both a sliding window of camera poses, as well as a fixed number of mapped landmarks [3], [18], or by separately building a map of the environment [19]. Although such methods bound the processing cost of SLAM (the number of mapped landmarks in the state vector is kept small), their performance during a hovering scenario hinges upon the criterion employed for selecting which features to be included in the state vector.…”

“…Such geometric constraints, provide corrections to the pose (position and orientation) estimates of an INS, and can significantly improve the localization accuracy within GPS-denied areas. As a result, recent advances in VINS have led to successful applications to ground [1], [2], aerial [3], [4], and space exploration [5] vehicles.…”

Towards Consistent Vision-Aided Inertial Navigation

“…Second, unlike laser-scanner-based methods that rely on the existence of structural planes [10] or height invariance in semi-structured en-vironments [30], using vision as an exteroceptive sensor enables VINS methods to work in unstructured areas such as collapsed buildings or outdoors. Furthermore, both cameras and IMUs are light-weight and have low power-consumption requirements, which has lead to recent advances in onboard estimation for Micro Aerial Vehicles (MAVs) (e.g., [36,37]). …”

“…However, these have focused on the simplified problem of estimating the 2D robot pose since the number of particles required is exponential in the size of the state vector. Existing work has addressed a variety of issues in VINS, such as reducing its computational cost [26,37], dealing with delayed measurements [36], increasing the accuracy of feature initialization and estimation [15], and improving the robustness to estimator initialization errors [21].…”