Machine Learning for Autonomous Jellyfish Orientation Determination and Tracking

Automated jellyfish classification and orientation determination can be used to improve the robustness of an autonomous jelly tracking system. A set of training images were segmented and augmented with orientation data, and a feature vector was generated for each image. An SVM was then trained to estimate a jelly’s orientation. In simulations on unseen test images, the approach was able to correctly determine the orientation of a jellyfish to within ±15◦ in 68% of cases. This is improved to within ±15◦ in 96% of cases on a training set. The algorithm was successful at categorizing the 3D angle of the jelly out of the image plane, and in certain cases was also able to distinguish between a jelly and a foreign object in the image frame.