Approximate Recursive Bayesian Filtering methods for robot visual search

Visual servoing is an essential enabling technology for robots operating in semi- and un-structured contexts, such as robot assistants working in collaboration with people. However, due to dynamic and unpredictable nature of such environments, existing methods of target tracking can lose visibility of task/target, leading to servo failure. In such situations, it is desirable that the robot reacquire the target in an autonomous/automatic fashion. In this paper we take a fresh look at this problem by examining the simplified case of a pan-tilt mounted camera visually searching for a lost target. We adopt Lost Target Search techniques based on Recursive Bayesian Filtering algorithms that have been applied to other search platforms such as aerial search and rescue. We investigated both an approximate grid-based filter and a sequential Monte Carlo method, namely particle filter. In both cases we use a new sensor-based observation model. The particle filter exhibited superior performance over approximate grid-based filter in our simulations, and was utilized in a follow-on experiment. In the experiment, we improved the particle filter performance by considering the a priori target tracking information in the motion model. Finally, we discuss the implications of this approach to higher degree of freedom robot systems.