Vision Planning for Object Search using Multiple Visual Features

This paper describes a vision planning method for object search using multiple visual features. To realize an efficient search, it is important to select appropriate search actions (i.e., features to use, fixation points, and resolution) depending on the status of the search. We first define similarity measures between the target object image and image regions in input images, and built similarity distribution models for the target object and the background. We then use the distribution for estimating the detection probability of the target in each location in the image for each search action. The planning method repeatedly selects the best search action which maximizes the expectation of the detection probability. We also develop a method of estimating the distribution models for the background from relatively simple image features. We compare the proposed method with methods with fixed search strategies to show its effectiveness.