Tracking multiple maneuvering point targets using multiple filter bank in infrared image sequence

Performance of any tracking algorithm depends upon the model selected to capture the target dynamics. In real world applications, no a priori knowledge about the target motion is available. Moreover, it could be a maneuvering target. The proposed method is able to track maneuvering or nonmaneuvering multiple point targets with large motion (∓20 pixels) using multiple filter bank in an IR image sequence in the presence of clutter and occlusion due to clouds. The use of multiple filters is not new, but the novel idea here is that it uses single-step decision logic to switch over between filters. Our approach does not use any a priori knowledge about maneuver parameters, nor does it exploit a parameterized nonlinear model for the target trajectories. This is in contrast to: (i) interacting multiple model (IMM) filtering which required the maneuver parameters, and (ii) extended Kalman filter (EKF) or unscented Kaiman filter (UKF), both of which require a parameterized model for the trajectories. We compared our approach for target tracking with IMM filtering using EKF and UKF for non-linear trajectory models. UKF uses the nonlinearity of the target model, where as a first order linearization is used in case of EKF. RMS for the predicted position error (RMS-PPE) obtained using our proposed methodology is significantly less in case of highly maneuvering target.