• No se han encontrado resultados

Since the VPS will be operating in a semi-structured environment it will be needing a full processing chain, including feature selection, neighbourhood scoring and con- straint validation in order to obtain an initial match. Once the first match has been correctly obtained the system can optimise itself by switching over to feature verifi- cation, which will significantly speed up the time to match and provide positioning data at a higher rate.

Descriptor

Given a landmark size of 80 x 80 pixels, a spacing of approximately 80 pixels between each landmark and a detection error of two pixels, the estimated angular error in the fingerprint is ± 1 degree. In order to match the database, each fingerprint encodes its ten nearest neighbours for validation.

Feature Selectors

As the environment is semi-structured, the VPS will use both the Gabriel Graph and Central feature selector. The Central feature selector is used to avoid matching

incomplete features near the edges of the frame, while the Gabriel graph eliminates poorly conditioned features such as features placed along a road.

The Central feature selector has been configured to use a linear scale, that provides a score from zero to one where zero is at the x or y limits of the image and one is at the centre of the image. As previously discussed in Section 5.1.1, this approach has been chosen as it favours more complete features in sparse scenes.

Meanwhile, the Gabriel graph requires very little configuration as it only requires a integer specifying the number of features to return. Thus the Gabriel graph algorithm will always return the same number of features but the scores for the features are normalised to the set, the first feature returned will always have a score of one, then the n next best features are returned to the matcher. In the UAV configuration the VPS will be time constrained, so the Gabriel graph has been configured to return the 20 best features in each scene.

Database Region Extraction

The system needs a number of parameters in order to extract the expected region from the database. This depends on the last known position, the vehicle’s velocity, control inputs from the autopilot and the drift on the inertial sensors. In order to simplify the problem the vehicle’s movement and control inputs have been taken out of the system. Once the simulated vehicle looses GPS it maintains the last known position with an increasing positioning error.

While it is possible to add the control inputs from the autopilot, these parameters have no meaningful effect on the performance of the descriptor and matcher, which are benchmarked on the recall rate (see Section 6.3 for further details). In addi- tion, the control inputs add further errors which cannot be modelled without more information about the vehicle and autopilot.

The region selection is thus computed as the last known position plus the estimated distance travelled according to the IMU. In addition, this region is expanded by the estimated drift in X and Y as a function of time seen in Figure 6.2. The drift has been scaled up by a factor of two to cover other unknown in the system, such as changes in altitude.

Thus, if a new image is captured dt seconds after the loss of GPS the search region is:

Xc = X0+ ∆XIM U (6.1)

where Xc is the centre of the search region, X0 is a 1D vector of the last known

position and ∆XIM U is the change in position integrated from the IMU. The size of

the search region is:

Xs = Sf + 2 ∗ D(dt) (6.2)

where Xs is a 1D vector composed of the width and height of the search region, Sf

is a vector of the sensor footprint and D(dt) is the drift error as a function of time since last GPS update.

Optimisation Threshold

If the estimated drift is low then the system can disable the initial candidate search. This threshold is dependent on the number of estimated additional features in the image, thus if the drift is low there will only be a small number of additional features, while a high drift error yields a large number of features and causes the direct matcher to fail. In this test the threshold has been set to 25 meters which, given the estimated feature density, gives up to 40 additional features in each frame depending on the type of scene.

Constraints

The system will only use the region similarity constraint since the search region is small and the last known position is known (the internal distance constraint will not be able to provide meaningful data when used with small search regions). Region similarity returns a score for the fit of each assignment and the highest scoring assignment is assumed to be correct.

When constraints are used for validation in the optimised matching method, the system returns the percentage of triangles that matched between the query and target set. The assignment is considered to be correct if at least 95% of the triangles match.