We currently use systematic sampling to sample every tenth observation in the Particle Filter to reduce overconfidence in the model. However, this biases the model quite badly towards minimizing error for close observations since laser scans are taken with angular sweeps (observation density is inversely proportional to the distance from an object).
I propose that we formulate some sort of density-based observation sampling to more evenly balance the error penalization contribution for distant objects and close objects.
We currently use systematic sampling to sample every tenth observation in the Particle Filter to reduce overconfidence in the model. However, this biases the model quite badly towards minimizing error for close observations since laser scans are taken with angular sweeps (observation density is inversely proportional to the distance from an object).
I propose that we formulate some sort of density-based observation sampling to more evenly balance the error penalization contribution for distant objects and close objects.