Three-dimensional forward-looking sonar target recognition with machine learning
AssigneeFarSounder, Inc.
InventorsMatthew Jason Zimmerman, Heath Henley, Austin Berard, Evan Lapisky
Life of this patent
Estimating the term…
Abstract
Machine learning algorithms can interpret three-dimensional sonar data to provide more precise and accurate determination of seafloor depths and in-water target detection and classification. The models apply architectures for interpreting volumetric data to three-dimensional forward-looking sonar data. A baseline set of training data is generated using traditional image and signal processing techniques, and used to train and evaluate a machine learning model, which is further improved by additional inputs to improve both seafloor and in-water target detection.
Claims21
A method for detecting and classifying underwater features, the method comprising:
- operating a three-dimensional forward-looking sonar (3D-FLS) sensor to obtain 3D-FLS data from the sensor, wherein the 3D-FLS data comprises a 3-dimensional point cloud of target strength values;
- providing the 3D-FLS data as input to a machine learning system that includes a neural network, wherein the neural network has been trained on training data comprising 3D-FLS training data comprising 3-dimensional point cloud of backscatter strengths and bathymetric survey data from a reference source;
- and
- using the machine learning system to detect a feature in the 3D-FLS data and classify the feature as seafloor or an in-water target.
A system for real-time detection and classification of underwater features, the system comprising:
- a three-dimensional forward-looking sonar (3D-FLS) device comprising a 3D-FLS sensor, wherein the 3D-FLS device is operable to insonify a region ahead of a vessel and collect 3D-FLS data from the 3D-FLS sensor, wherein the 3D-FLS data comprises a 3-dimensional point cloud of target strength values;
- and
- a computing system coupled to the 3D-FLS device, the computing system comprising a neural network, wherein the neural network has been trained on:
- (i) labeled 3D-FLS training data comprising 3-dimensional point clouds of backscatter strengths;
- and (ii) bathymetric survey data from a reference source, such that the neural network is operable to detect features in the 3D-FLS data and classify the features as seafloor or in-water targets.
Dependencies are inferred from the claim text.