Real-time anomaly detection in side-scan sonar imagery for adaptive AUV missions

Abstract

Autonomous Underwater Vehicle (AUV) operations are inherently bandwidth limited but increasingly data intensive. This leads to large latencies between the capture of image data and the time at which operators are able to make informed decisions using the results of a survey. As AUV endurance and reliability continue to improve, there is a greater need for real-time data processing to inform on-board adaptive mission planning. In this paper, we present an anomaly detection framework based on saliency and rarity and demonstrate it using existing side-scan sonar datasets collected by an AUV. Salient regions are first identified using a novel method with analogies to keypoint detection in traditional image processing. Models of these regions are then learned to determine rarity using an online approach for real-time use during a mission. The algorithm we present will be implemented in field trials later this year. This approach to adaptive mission planning enables an AUV to both resurvey anomalies at higher resolutions and selectively transmit imagery for operator analysis and feedback within the scope of a single deployment.