Change detection for undermodelled processes using mismatched hidden Markov model test filters

In this paper we present a change detection approach for dependent processes based on the output of a mismatched hidden Markov model (HMM) test filter (i.e., a HMM filter applied to observations not generated by its model). The presented approach is intended to be suitable for dependent processes that are significantly undermodelled in the sense that their conditional densities are not known, are too complex, or are otherwise unsuitable for existing change detection techniques. We establish a description of a mismatched HMM test filter's output when it is applied to sequences generated by a general dependent process. This description is used to motivate the proposal of a novel change detection approach based on monitoring the statistical properties of the mismatched HMM test filter's output. We examine our proposed approach in an important vision based aircraft detection application where it offers improvements in detection range (mean increase of 276m) compared to the current state of the art baseline detection approach