Detection and tracking of divers for underwater human-robot interaction scenarios

Abstract

The underwater domain is a dangerous and complex environment for human divers. Often, divers have to monitor their own life support systems as they navigate to the work site or operate dangerous machinery. Military divers have to navigate for extended periods of time without surfacing or without using localization techniques that might give away their positions. Human divers have operated under these harsh conditions for decades with few advancements in technology. In fact, a diver performs the basic task of navigation by aligning the body with a compass and counting leg kicks (i.e., human-oriented dead-reckoning). It is proposed that an Underwater Robotic Assistant (UWRA) will improve the efficiency and safety of the diver's underwater operations by providing several key capabilities. For example, the UWRA can provide navigation assistance, ferry tools from the surface, enter structures too dangerous for human divers, and carry hazardous materials. However, in unstructured environments, underwater robots are limited in their ability to localize and track a human diver at the resolution required to enable diver-robot interactions. Optical cameras can be rendered useless by the turbidity of the water, localizing radio signals do not propagate well through the water medium, and acoustic positioning systems can be expensive to deploy. We propose that by developing novel 2D imaging sonar processing techniques, an underwater robot can detect, track, and trail a human diver. The objective of this research is to detect and track human divers in 2D imaging sonar data. While the physical properties of sonar allow it to detect objects at longer ranges than optical cameras in underwater scenarios, it is plagued with noise and multi-path propagation. Also, when a diver is ensonified with a 2D imaging sonar, a fragmented acoustic reflection is returned. The fact that a single object can produce multiple returns means that tracking the human diver cannot be solved by applying traditional multiple hypothesis tracking algorithms, which operate on the assumption of each object generating only a single measurement. To overcome the sonar noise and multiple fragmented returns, we developed a novel adaptive thresholding algorithm and a hierarchical multiple object tracking algorithm. While the Kalman filter is extensively used in our tracking algorithms, we developed a novel method for adaptively modifying the Kalman filter's measurement matrix to track objects that generate multiple measurements.