Data-driven transition matrix estimation in probabilistic learning models for autonomous driving

Abstract

A novel approach is presented for learning probabilistic transition matrices from temporal data series as switching models based on Generalized States (GS). An observed data sequence is analyzed by a reference filter whose errors are clustered. Each cluster is associated with a dynamic flow model, which described as a parametric linear attractor. The set of linear attractors define the Hierarchical Generalized Dynamic Bayesian Network (H-GDBN), which encodes a learned model of the obtained sequence. A Markov Jump Particle Filter (MJPF) uses H-GDBN’s probabilistic information to make inferences at a multilevel scale and facilitates the detection of abnormalities. This paper shows how transition matrices can be obtained as an integral part of the clustering step by employing the advantages of GSs, enabling a unique optimal criterion for learning flow models at discrete and continuous levels of H-GDBN. For evaluating the proposed method, odometry and proprioceptive control data from an autonomous vehicle are employed to learn H-GDBNs. Learned H-GDBN are used to detect abnormalities when vehicle encounter any abnormal situation. Performance evaluation based on ROC curves is provided to select the optimal transition matrix.