Hybrid Belief Pruning with Guarantees for Viewpoint-Dependent Semantic SLAM

Semantic simultaneous localization and mapping is a subject of great interest in robotics and AI that directly influences the autonomous vehicles industry, army industries, and more. One of the main challenges in this field is to obtain object classification jointly with robot trajectory estimation and environment mapping. Considering the coupling between these random continuous and discrete random variables, one has to maintain a hybrid belief, where the number of class hypotheses raises exponentially by the number of objects. A common solution is to prune hypotheses with a sufficiently low probability and thus maintain only a small fraction of the hypotheses. However, after pruning and renormalization, the resulting hypotheses’ probabilities are overconfident with respect to the original probabilities without pruning. This is especially problematic when the robot decides on which action to do according to the objects’ classes and classification. In this framework, we present a method to maintain more accurate probabilities of the objects’ classes, under the assumption that the position of the robot and the objects are known. Our method provides a guarantee on the belief after pruning in the form of a lower bound on the probabilities of the maintained hypotheses. It runs in an incremental manner and with high efficiency. This method is particularly useful for highly reliable systems, where an overconfident classifier without guarantees may not be suitable.