engleski

Moving objects detection and tracking by omnidirectional sensors of a mobile robot

Directional data emerge often in many aspects of mobile robotics. Measurements from various sensors yield direction-only information of the objects of interest. Since probabilistic methods have been widely accepted and successfully utilized in many mobile robotics problems, question arises if such modeling could offer prospects in the context of probabilistic representation of directional data gathered by a mobile robot. One of the goals of this thesis is to develop directional statistics based methods for moving object tracking by omnidirectional sensors of a mobile robot. In that mindset the thesis addresses moving object tracking via two different problems, namely speaker detection, localization and tracking with a microphone array, and moving object detection, tracking and following with an omnidirectional camera. Furthermore, in the thesis we also address the challenge of heterogeneous sensor fusion through the prism of moving object tracking. The speaker localization and tracking problem is solved by modeling the measurement of a microphone array with a convex mixture of von Mises distributions, where the tracking is thus performed by way of particle filtering. This approach is later extended, to circumvent the sample based techniques, by keeping the tracking procedure fully in the analytical domain via a mixture filter based on the von Mises distribution. Furthermore, a prerequisite for robust speaker localization and tracking is voice activity detection. In the thesis we analyze this problem from the standpoint of model based voice activity detection methods which are enhanced by supervised learning algorithms. Specifically, a detector based on the Rayleigh and Rice distributions is coupled with a number of carefully chosen spectral and temporal features in a supervised classification approach. Apropos of omnidirectional camera, where spherical projection model coupled with displacement information from motor encoders is proposed to segment out features that do not belong to the static scene around the mobile robot, directional statistics is used in the context of movement tracking on the sphere with a Bayesian filter based on the von Mises-Fisher distribution. Finally, fusion of heterogeneous sensors for object tracking is analyzed in a comparative study of the extended information filter, the unscented information filter and the particle filter.

moving object tracking; directional statistics; speaker localization; von Mises distribution; microphone array; voice activity detection; sensor fusion; mobile robotics

nije evidentirano