Real-Time Video Tracking Using Convolution HMMs
Javier R. Movellan, John R. Hershey, Josh Susskind · 2004
Bayesian filtering provides a principled approach for a variety of problems in machine perception and robotics. Current filtering methods work with analog hypothesis spaces and find approximate solutions to the resulting non-linear filtering problem using Monte-Carlo approximations (i.e., particle filters) or linear approximations (e.g., extended Kalman filter). Instead, in this paper we propose digitizing the hypothesis space into a large number, n 100, 000, of discrete hypotheses. Thus the approach becomes equivalent to standard hidden Markov models (HMM) except for the fact that we use a very large number of states. One reason this approach has not been tried in the past is that the standard forward filtering equations for discrete HMMs require operations per time step and thus rapidly become prohibitive. In our model, however, the states are arranged in two-dimensional topologies, with locationindependent dynamics. With this arrangement predictive distributions can be computed via convolutions. In addition, the computation of log-likelihood ratios can also be performed via convolutions. We describe algorithms that solve the filtering equations, performing this convolution for a special class of transition kernels in order n operations per time step. This allows exact solution of filtering problems in real time with hundreds of thousands of discrete hypotheses. We found this number of hypotheses su#cient for object tracking problems. We also propose principled methods to adapt the model parameters in non-stationary environments and to detect and recover from tracking errors.