Anomaly Detection in Multibeam Echosounder Seabed Scanssipl.eelabs.technion.ac.il/wp-content/uploads/sites/6/... · 2017. 7. 12. · Multibeam Scans •Echo sounding multibeam scanner.

Ron Vainshtein

Itay Cohen

Supervisors:

Dr. Alon Amar, Yaakov Buchris

In Collaboration with: Azriel Sinai

Anomaly Detection in Multibeam Echosounder Seabed Scans

Motivation

• Multibeam seabed data.

• Anomaly detection using scans.

• Anomalies:• Mines• Pipes • Wreckage • Waste

2/35

Multibeam Scans

• Echo sounding multibeam scanner.• SeaBat T20P by Teledyne-Reson

• 50 pings per second, 512 beams per ping

• Seabed mapping via beam echo.

• Scans affected by:• Wind

• Waves

• Vessel movement

• Depth3/35

Scan Characteristics

• Low resolution point cloud.

• Scan lines pattern.

• Spacing, direction and depth vary.

• Sparse and not sampled uniformly.

• Missing sample batches.4/35

Challenges

• Scan analysis and processing is difficult.

• Little prior information about targets.

• Small scan dataset.

• No prior work.

Anomaly detection is challenging!

5/35

Goals

•Anomaly detection for multibeam seabed scans.

•Overcome difficulties arising from the data.

6/35

Scan Properties in Detail

shore

7/35

• Lines are spaced 0.2 - 0.7m.• Depends on vessel orientation.

Line Spacing in Deep Water

9/35

Targets in Deep Water

11/35

Missing Data

• Beams scattered away from scanner.• Scanner discards unreliable samples.

scanner

12/35

Algorithm Stages

Holes Detection

Filling Holes

Anomaly Detection

Multibeam Scan

Regions of Interest

Holes Detection

13/35

Delaunay Tessellation Field Estimator• Used in cosmology.

• Local density calculation.

• Adaptive to variations in density and geometry.

Holes Detection

[Schaap & Van De Weygaert ’00]

14/35

Delaunay Triangulation

• Circumscribed circle contains only the inscribed triangle points.

• Guarantees immediate neighbors.

• Efficient O(nlog(n)).

Holes Detection

15/35

DTFE – Density Calculation

• Take sum of areas of participating triangles.

• Density at each vertex = 1

sum.

• Linearly interpolate density.

Holes Detection

16/35

DTFE-Based Hole Contouring

• Apply DTFE and assign density to triangles.

• Low density triangles indicate holes.

• Create connected components from triangles.

Holes Detection

17/35

DTFE-Based Bole Contouring

• Use threshold on area to avoid small holes.

• Defines explicitly where to fix the data.

Holes Detection

18/35

Algorithm Stages

Holes Detection

Filling Holes

Anomaly Detection

Multibeam Scan

Regions of Interest

19/35

Filling Holes

Defining the sampling points:• Sampling grid based on the triangulation and data.

Choosing the Interpolation method:• Standard interpolation methods (linear, cubic, etc.) - poor reconstruction.

• A multiscale iterative approach is used.

Filling Holes

20/35

Interpolation Method

Laplacian Pyramid Extension

• Pyramids – multiscale image manipulation.

• Create stack of increasing scales from surroundings.• No downsampling is used.• Each scale is created using all previous scales.

• Suitable for scattered data.

Filling Holes

[Bermanis, Coifman and Averbouch. ’13]

21/35

Stack GenerationFilling Holes

22/35

Filling Holes - ResultsFilling Holes

23/35

Filling Holes – Deep WaterFilling Holes

24/35

Algorithm Stages

Holes Detection

Filling Holes

Anomaly Detection

Multibeam Scan

Regions of Interest

25/35

ROI Detection

>Threshold?

26/35

ROI Detection

• Variance map is calculated with several window sizes.

• ROI decision – voting process on layers in patches.

• Patch size and threshold are parameters.

Anomaly DetectionAnomaly Detection

27/35

Results – Target Anomaly Detection

28/35

Results - Target

ROI 1.6m

Anomaly Detection

29/35

Results - Background Anomaly Detection

30/35

Results - Background

ROI 1.6m

Anomaly Detection

31/35

Summary

• Multibeam data is challenging

• Algorithms from various fields

• New interpolation framework

• Innovative anomaly detection method

32/35

• Will be used in new naval products.

• Fits several seabed-related applications.

• Publish a paper.

• Develop anomaly classification algorithm.

Achievements and Future Work

34/35

Acknowledgments

• Yaakov Buchris

• Dr. Alon Amar

• SIPL staff:• Nimrod Peleg

• Yair Moshe

• Ori Bryt

• RAFAEL: Azriel Sinai

35/35