Density Tree Estimator in Action

This video shows a target tracking application using an Octree. It localizes and tracks a target using a single omni-directional microphone. Setup: Target has 3-dimensional state (Octree): - (x,y) - position - (alpha) - movement direction The target has one sensor capable of measuring the distance to the next wall without knowing which wall or the direction of the measurement. The measurement is heavily disturbed by Gaussian zero mean noise. Furthermore the target has a map of its surroundings (shown in gray in the video). ------------------------------------------------------------------------ Video: The video shows three marginals of the estimated system state density: - (x,y) - lower left - (x,alpha) - upper left - (y,alpha) - lower right - likelihood(x,y) - upper right The black pole denotes the true robot state. ------------------------------------------------------------------------ Experiments: 1. Wake-up robot: There is no initial knowledge of the targets system state. The target moves around the central pole in a circle making 20 measurements. Finally it has localized itself resulting in a bimodal density for the system state. 2. Kidnapped robot: Using the resulting density from the first experiment the robot is "kidnapped". Thus, its orientation and position is changed randomly without updating the system state density. Then target moves around the central pole in the opposite direction twice, while making 40 measurements, before it has a robust estimate of its position and orientation. The number of nodes in the tree is roughly kept at 20000 during the experiments. ---------------------------------------------------------- This animation was generated using the Povray_Matlab software developed at the ISAS-Lab. Further information: http://isas.uka.de/en/User:Eberhardt