Roberts, J.M. and Charnley, D.
Multiple Model Filtering for Time-to-Collision Estimation and Segmentation
IEEE Trans. on Signal Processing
Full text not available from this repository.
This paper addresses the problem of time-to-collision/recession-rate estimation of tracked two-dimensional image features over a long image sequence using the Multiple Model Adaptive Estimator (MMAE). Extended Kalman filters are constructed assuming that an image feature moves on a constant plane with respect to an observer, and simulations are then presented. These simulation results show that a filter based on a single motion model is not appropriate when trying to estimate the time-to-collision of a tracked feature point in a typical driving scenario, due to either undermodelling or overmodelling of the expected feature motion. A Multiple Model Filter algorithm (the MMAE2) is then proposed and investigated to overcome these problems and to help with scene segmentation. This algorithm was found to be unreliable when tested on simple artificially generated feature trajectories and could not discriminate reliably between models of different orders. Finally, a modified version of the MMAE2 using empirically generated estimates of the residual's covariance, instead of the theoretical covariance traditionally used, is proposed. The modified algorithm (MMAE2) was tested and was found to discriminate reliably between models and therefore produce accurate estimates of recession-rate.
Actions (login required)