A Probabilistic Approach for Fusing People Detectors

Abstract

Automatic detection of people is essential for automated systems that interact with persons and perform complex tasks in an environment with humans. To detect people efficiently, in this article it is proposed the use of high-level information from several people detectors, which are combined using probabilistic techniques. The detectors rely on information from one or more sensors, such as cameras and laser rangefinders. The detectors’ combination allows the prediction of the position of the persons inside the sensors’ fields of view and, in some situations, outside them. Also, the fusion of the detector’s output can make people detection more robust to failures and occlusions, yielding in more accurate and complete information than the one given by a single detector. The methodology presented in this paper is based on a recursive Bayes filter, whose prediction and update models are specified in function of the detectors used. Experiments were executed with a mobile robot that collects real data in a dynamic environment, which, in our methodology, is represented by a local semantic grid that combines three different people detectors. Results indicate the improvements brought by the approach in relation to a single detector alone.