Automated open circuit scuba diver detection with low cost passive sonar and machine learning

Abstract

This thesis evaluates automated open-circuit scuba diver detection using low-cost passive sonar and machine learning. Previous automated passive sonar scuba diver detection systems required matching the frequency of diver breathing transients to that of an assumed diver breathing frequency. Earlier work required prior knowledge of both the number of divers and their breathing rate. Here an image processing approach is used for automated diver detection by implementing a deep convolutional neural network. Image processing was chosen because it is a proven method for sonar classification by trained human operators. The system described here is able to detect a scuba diver from a single acoustic emission from the diver. Twenty dives were conducted in support of this work at the WHOI pier from October 2018 to February 2019. The system, when compared to a trained human operator, correctly classified approximately 93% of the data. When sequential processing techniques were applied, system accuracy rose to 97%. This demonstrated that a combination of low-cost, passive sonar and a properly tuned convolutional neural network can detect divers in a noisy environment to a range of at least 12.49 m (50 feet).