A filter method for pose estimation of maneuvering target

A filter method is presented for 3D pose (position and orientation) estimation of an arbitrary moving target from a monocular image sequence. Problem considered here is error propagation from image features to pose parameters. To this end, we have derived a combination of two filter schemes. The first filter scheme uses maneuver detection technique, in which the criterion of an optimal detector is deduced, two detectors suitable for fast and slow maneuvers are analyzed, and limited memory filtering is adopted for maneuver correction. In the second filter, a robust model is established via Lagrange interpolation and numerical integration, and the possibility of filter divergence is avoided by employing an adaptively estimated fading factor. Finally, generalized pseudo Bayes algorithm is employed to combine the two filter models for better filter performance. Superior to previous approaches that were limited to slow and smooth motion, this filter method is applicable for a maneuvering target that acts in an unknown manner. Moreover, it is suitable for real-time environment in terms of the following aspects: the structure of the filter is simplified by avoiding the need of EKF, the computational cost is much reduced by running six filters on the six pose sequences in parallel; the process to extract the feature points is simplified due to the predicted feature locations; time efficiency is increased because the sequence of image frames is allowed to be at unequal intervals. Simulation results and implement of a system for multi-mobile robots formation show the capacity of this algorithm.