Search

This paper describes the Random Finite Set approach to Bayesian mobile robotics, which is based on a natural multi-object filtering framework, making it well suited to both single and swarm-based mobile robotic applications. By modeling the measurements and feature map as random finite sets (RFSs), joint estimates the number and location of the objects (features) in the map can be generated. In addition, it is shown how the path of each robot can be estimated if required. The framework differs dramatically from existing approaches since both data association and feature management routines are integrated into a single recursion. This makes the framework well suited to multi-robot scenarios due to the ease of fusing multiple map estimates from swarm members, as well as mapping robustness in the presence of other mobile robots which may induce false map measurements. An overview of developments thus far is presented, with implementations demonstrating the merits of the framework on simulated and experimental datasets.

Most visual multi-target tracking techniques in the literature employ a detection routine to map the image data to point measurements that are usually further processed by a filter. In this paper, we present a visual tracking technique based on a multi-target filtering algorithm that operates directly on the image observations and does not require any detection nor training patterns. Instead, we use the recent history of image data for non-parametric background subtraction and apply an efficient multi-target filtering technique, known as the multi-Bernoulli filter, on the resulting grey scale image data. In our experiments, we applied our method to track multiple people in three video sequences from the CAVIAR dataset. The results show that our method can automatically track multiple interacting targets and quickly finds targets entering or leaving the scene.