Particle filtering strategies for data fusion dedicated to visual tracking from a mobile robot

This paper introduces data fusion strategies within particle filtering in order to track people from a single camera mounted on a mobile robot in a human environment. Various visual cues are described, relying on color, shape or motion, together with several filtering strategies taking into account all or parts of these measurements in their importance and/or measurement functions. A preliminary evaluation enables the selection of the most meaningful visual cues associations in terms of discriminative power, robustness to artifacts and time consumption. The depicted filtering strategies are then evaluated in order to check which people trackers regarding visual cues and algorithms associations best fulfill the requirements of the considered scenarios. The performances are compared through some quantitative and qualitative evaluations. Some associations of filtering strategies and visual cues show a significant increase in the tracking robustness and precision. Future works are finally discussed.