Automatic target recognition in sonar imagery using a cascade of boosted classifiers

Abstract

This thesis is concerned with the problem of automating the interpretation of data representing the underwater environment retrieved from sensors. This is an important task which potentially allows underwater robots to become completely autonomous, keeping humans out of harm’s way and reducing the operational time and cost of many underwater applications. Typical applications include unexploded ordnance clearance, ship/plane wreck hunting (e.g. Malaysia Airlines flight MH370), and oilfield inspection (e.g. Deepwater Horizon disaster). Two attributes of the processing are crucial if automated interpretation is to be successful. First, computational efficiency is required to allow real-time analysis to be performed on-board robots with limited resources. Second, detection accuracy comparable to human experts is required in order to replace them. Approaches in the open literature do not appear capable of achieving these requirements and this therefore has become the objective of this thesis. This thesis proposes a novel approach capable of recognizing targets in sonar data extremely rapidly with a low number of false alarms. The approach was originally developed for face detection in video, and it is applied to sonar data here for the first time. Aside from the application, the main contribution of this thesis, therefore, is in the way this approach is extended to reduce its training time and improve its detection accuracy. Results obtained on large sets of real sonar data on a variety of challenging terrains are presented to show the discriminative power of the proposed approach. In real field trials, the proposed approach was capable of processing sonar data real-time on-board underwater robots. In direct comparison with human experts, the proposed approach offers 40% reduction in the number of false alarms.