Automated matching and pre-classification of sonar tracks on known navigational paths

When utilising high-resolution broadband sonars in littoral waters, the sonar image can become highly cluttered with false alarms and tracks. These false alarms usually originate from a myriad of different environmental sources, such as seamounts, underwater ridges, and other topographical features, or man-made objects, such as pipelines. Automatic classification algorithms may be developed through supervised learning and be employed for reducing the amounts of false alarms. However, this requires large amounts of labelled data. This is particularly the case for methods using deep neural networks. Labelling is a tedious and time consuming endeavour. Here we propose a method for automatic recognition of sonar tracks on known navigational paths by utilising known parameters, such as bearing, range, course, and velocity, from both the navigational and sonar track. In this work we present the method, perform a parameter study and a sensitivity analysis. Due to the limited amount of sonar data with corresponding known navigational data from possible target vessels, and the minimal maneuvering usually performed by such vessels, the sonar data used in this study is synthetic containing several synthetic target vessels. We find that the method is robust and can correctly match up to 84% of tracks on a navigational path while discarding 99.5% of false tracks. Although the ratio of false to target tracks in the classified data set should ideally be even lower for supervised learning, the large reduction in false tracks will ease the process of manually removing the remaining 0.5% drastically.