Finding Tiny People: Long-range outdoor sensing for establishing joint attention in human-robot interaction

In this thesis, we present two novel methods for a robot to determine whether a distant human wants to interact. The first method finds periodic signals in camera streams and clusters them into potential human waving gestures. We demonstrated this on a ground robot by waving at the robot from long distances of up to 45 meters to attract its attention. The robot then approached to close range to confirm the gesture with other, more precise but distance-limited methods, such as human detection. The second contribution is a robot behaviour for establishing joint attention with a human by exploiting trajectory signals. If a human is detected on an intercept course with the robot, the robot can vary its trajectory to probe the intent of the human. If the human corrects its trajectory to maintain the intercept, this can be considered a strong signal that the human wants to interact with the robot. We show that under modest assumptions, an arbitrary level of confidence in human intent can be achieved by iterating the behaviour. In addition to contributions to human-robot interaction, we present a novel outdoor robot dataset captured at Simon Fraser University campus. The dataset consists of hundreds of gigabytes of trail data recorded using a ground-based robot equipped with six cameras and two laser scanners.