Concepts de base
Terrain characterization techniques can improve the performance of automated sonar processing algorithms by adapting to the complexity of the underwater environment.
Résumé
The paper presents two methods for terrain characterization to enhance the performance of Automated Target Recognition (ATR) algorithms on side-scan sonar imagery:
-
Algorithm-led Terrain Characterization:
- Quantitatively measures terrain complexity by simulating the insertion of contacts and evaluating the ATR algorithm's performance.
- Provides an objective, application-driven metric of terrain complexity based on the probability of detection and false alarm density.
-
Operator-led Terrain Classification:
- Employs an unsupervised machine learning approach to cluster different seafloor types.
- Incorporates subject matter expertise by allowing operators to label the clusters, enabling contextual and explainable terrain characterization.
- Produces a multi-class terrain classification map that aligns with the human understanding of seafloor complexity.
Both methods are designed for real-time, onboard processing on Autonomous Underwater Vehicles (AUVs) with minimal human intervention, making them suitable for operational deployment.
The terrain characterization information can be used by the AUV's autonomy framework to adapt the sensor configurations and algorithms, improving mine detection rates and reducing false alarms. It can also be used to optimize mission planning by identifying obstacles and prioritizing search and clearance efforts.
The paper discusses lessons learned, such as the need for a 2D representation of terrain complexity to account for viewpoint dependence, and the potential use of bathymetry data to infer terrain characteristics without the need for multiple sonar passes.
Stats
The performance of Automated Target Recognition (ATR) algorithms on side-scan sonar imagery can degrade rapidly when deployed in non-benign environments.
Complex seafloors and acoustic artifacts can create false detections or prevent the detection of true objects.
Citations
"The performance of detection algorithms is dependent on the environment as various environmental factors can increase the difficulty of dissociating an object from background data."
"Terrain complexity characterisation enables mission prioritisation by identifying areas of interest and optimizing search and clearance efforts during the mission."