Mixed-Domain Periodic Motion (MDPM) tracker introduces a robust and efficient algorithm for an underwater robot to detect and track its companion diver. It uses both spatial domain and frequency domain features to track human swimming motion in spatiotemporal volume. The motion direction of a diver is modeled as a sequence of non-overlapping image regions over time, and it is quantified by the corresponding vector of intensity values. A Hidden Markov Model (HMM)-based pruning method exploits these intensity values to track a set of promising motion directions. The potentially optimal motion directions are then validated based on their frequency domain signatures, i.e., high energy responses in the 1-2 Hz frequency bands.
Experimental evaluation of the proposed MDPM tracker is performed using video footage recorded from the camera of an underwater robot with multiple divers swimming in different directions. Used datasets include both open-water and closed-water environments having a significant difference in lighting conditions and visibility. For further results and details, please check out this paper.