Data fusion for person identification in people tracking

This article presents four different approaches for identifying persons during tracking. The ability to keep track of people is essential for mobile robots which provide services to humans. The tracking task is usually carried out using range measuring devices like laser range sensors, which provide spatial information only. They allow to estimate the trajectories of objects quite accurately, but they do not allow to directly distinguish between the objects being observed. The tracking algorithms, therefore, tend to confuse tracks in difficult situations, like persons walking in close proximity to each other. The techniques presented in this article mitigate this problem by fusing the spatial measurements with different types of complimentary sensor data. The first technique uses colour information extracted from camera images to distinguish between persons, based on the colour of the clothes they wear. The second technique uses reflectance intensities, which can directly be provided by the laser range sensors for this purpose. The third method is again based on camera information, but it employs a probabilistic shape and motion model for each person it tracks, in order to distinguish between them. Finally, we present an approach that uses a network of dedicated sensors, which directly transmit identification information for each person. The aim of this paper is to give a comparative overview of these four methods, and to discuss their drawbacks and advantages.