Dynamic measurement clustering to aid real time tracking

Full-3D Edge Tracking with a Particle Filter

“…Even if the number of particles is raised to 500, tracking remains fragile and tends to fail. The poor performance is presumably partially due to the camera motion, which is smooth but rapid (and thus suited for a constant velocity model rather than the one used here); further, [6] uses a texture change-point edge detector rather than the high intensity gradient model used here.…”

“…A further comparative test was performed on a sequence of the maze model tracked successfully in [6]. This difficult sequence in a highly textured environment with many parallel edges was not trackable at 30Hz.…”

“…A disadvantage of the proposed system is increased jitter in stationary scenes. Further, the simple likelihood model dictated by real-time constraints makes the integration of more advanced edge detection [16,6] difficult.…”

“…[1] achieves this using RANSAC to estimate individual edges and case deletion to estimate pose. [6] goes beyond this by constructing a multi-modal likelihood estimate from RANSAC at the low level and an exhaustive search at high level. A common alternative to RANSAC is the use of a robust M-estimator [3,9,7,17].…”

“…Approaches to increase frame-to-frame robustness have included the use of robust estimation techniques [1,3], the addition of external sensors [7], and the use of point features for initialisation [17,15]; Further, attempts at considering multiple edge correspondence hypotheses have shown great potential [6]. However, all of these methods are fundamentally uni-modal in that they calculate a single (two for [15]) Gaussian posterior pose for each frame, which then provides a single prior pose for the next frame; if the estimate is sufficiently incorrect, tracking will fail.…”