Integrating multiple deep learning models to track and classify at-risk fish species near commercial infrastructure

Companies must not harm species at risk around their fixed infrastructure and need a way to detect and monitor at risk fish. However, a species at risk cannot be tagged and studied using conventional surgically implanted fish tracking technology. Innovasea is therefore developing a platform to monitor fish using a combination of sensors such as acoustic devices, visual and active sonar and optical cameras. This effort requires a robust accurate method to detect fish and classify them by species. To meet this need, we propose testing two deep learning models, one using filtering combined with a convolutional neural network for detection and the other using optical flow and gaussian mixture models with a YOLO fish species classifier. We will first test these methods for classifying fish species using data from a high resolution DIDSON imaging sonar, then develop guidelines to help Innovasea develop and launch their integrated multi-sensor platform for fish monitoring, and finally we will develop the best performing prototype model into a fully implemented system which will be able to detect and classify untagged fish with data from multiple types of sensors.

Faculty Supervisor:

Christopher Whidden;Luis Torgo

Student:

Vishnu Vardhan Kandimalla

Partner:

InnovaSea Marine Systems Canada Inc

Discipline:

Computer science

Sector:

Manufacturing

University:

Dalhousie University