Multi-sonar adaptive target detection using the Sphericity Test

Detection of underwater objects in sonar imagery is highly challenging task. Human operators often times outperform computer algorithms in moderate and difficult environments due to the high variability in environmental conditions and human proficiency in incorporating in-situ information and context. Most automatic target recognition algorithms are pre-trained in a different environment leading to high levels of model mismatch when new environments are encountered. This paper presents an environmentally adaptive algorithm for underwater target detection in multiple sonar images using the Sphericity Test. By estimating local second-order statistics within the image, the proposed application of the Sphericity Test is capable of in-situ adaptation in changing underwater environments. The proposed detector is applied to a dataset consisting of pairs of beamformed images collected over multiple target fields. The method will also be compared and benchmarked to a similar method which does not employ environmental adaptation. Performance measures including probability of detection/false alarm and Receiver-Operator Characteristic (ROC) curve metrics will be used to evaluate the performance of each detection method.