Method for detecting and recognizing boats

1. A computer implemented method for recognizing boats in a port, each boat having a bow, the method comprising:

receiving, by a receiver, a digital image of the port that includes one or more boats

establishing, by one or more processor, land/water boundaries from the digital image of the port including one or more boats, using a land mask and image statistics, without requiring a site model;

utilizing, by the one or more processor, the established land/water boundaries to determine thresholds and normalization parameters for the one or more boats;

detecting, by the one or more processor, one or more bow candidates for the boats in the image responsive to the determined thresholds and normalization parameters;

computing, by the one or more processor, an orientation of each of the boats with detected bows;

computing first features along the bow candidates to eliminate false bow candidates responsive to the orientations;

fitting, by the one or more processor, boat models to the detected one or more bow candidates, responsive to the orientations and the first features;

determining, by the one or more processor, one or more best fit models from the fitted boat models;

calculating second features in an interior, exterior and outline of the best fit models;

applying, by the one or more processor, a statistical classifier to score each of the one or more best fit models and to reduce false alarms for the best fit models; and

eliminating, by the one or more processor, lower scoring best fit models whose outline overlaps an outline of a higher scoring best fit model to obtain a more accurate result for the recognizing boats.

2. The method of claim 1 , wherein boat models are 2-dimensional or 3-dimensional models stored in a database.

3. The method of claim 1 , further comprising:

sorting, by the one or more processor, the remaining best fit models by their scores in a list;

determining, by the one or more processor, whether any best fit model with a lower score has overlap with a best fit model with a higher score; and

eliminating said best fit model with a lower score from the list.

4. The method of claim 1 , wherein computing first features along the bow candidates further comprises:

initiating, by the one or more processor, a bow search for each pixel the gradient magnitude of which, is above a determined threshold;

computing an initial bow orientation using a multi-pixel template at regularly spaced orientations looking for an orientation that has the highest contrast with an immediate background in front of and to the sides of the bow;

refining, by the one or more processor, the bow orientation using a weighted histogram of a bow edge phase on each side of the bow;

extracting, by the one or more processor, features along the bow edges and from an immediate background in front of the bow;

utilizing the computed initial bow orientation, stepping along a front portion of the boat candidate looking for the best pairs of boat edge pixels on each side using gradient magnitude times the pixel's gradient phase match against an expected phase computed from the initial bow orientation;

selecting, by the one or more processor, pairs of pixels whose metrics pass a determined threshold;

for the selected pairs of pixels, computing an average gradient magnitude of each side and an absolute difference between a distance from one side of the bow to a center line and a distance between the other side and the center line;

computing an angle difference between a phase at the point on one side with the bow orientation direction and the same metric for the other side and calculating an absolute difference of the angle difference to calculate features;

using, by the one or more processor, the calculated features to compute a bow score; and

eliminating, by the one or more processor, bows with bow scores lower than a predetermined value.

5. The method of claim 4 , further comprising applying by the one or more processor, thresholds to the bow score and to the calculated features to determine which bow candidates to eliminate.

6. The method of claim 4 , wherein the bow score is computed by multiplying the calculated features together.

7. The method of claim 6 , wherein the calculated features together include one or more of a number of bow pairs, an average gradient magnitude over all the bow points, an average side angle difference over all the bow pairs, an average width difference between the distance from one point of the bow pair to a center and a distance of the other point to the center; and an average gradient phase match of all the bow points to an expected phase.

8. The method of claim 1 , wherein fitting boat models to the detected bow candidates further comprises:

for a particular offset, orientation and model, computing, by the one or more processor, an average gradient magnitude and an average gradient phase-match between a gradient phase of the image at the outline points and an expected phase from the model; and

computing a model score as a function of a score for bow candidate, the average gradient magnitude and the average phase-match from the model outline points; and

determining, by the one or more processor, the best fit boat models for each bow candidate.