Keywords

relative pose estimation, essential matrix, unmanned aerial vehicle, multiple target tracking, motion detection, parallax

Abstract

We present ReSORtSAC: Recursively-seeded optimization, refinement, sample, and consensus. ReSORtSAC is a novel algorithm that can be used to estimate the relative pose between consecutive frames of a video sequence. Relative pose estimation algorithms typically generate a large number of hypotheses from minimum subsets and score them in order to be robust to noise and outliers. The relative pose is often represented using the essential matrix. Previous methods calculate essential matrix hypotheses directly without utilizing prior information. These equations are complex to evaluate and can return up to ten essential matrix solutions for each minimum subset, all of which must be scored.

Instead, we calculate relative pose hypotheses by optimizing the rotation and translation between frames, rather than calculating the essential matrix directly. The equations used in our optimization are simpler to evaluate, resulting in faster computation speeds. We also reuse the best hypothesis to seed the optimizer which reduces the number of relative pose hypotheses which must be generated and scored. These advantages greatly speed up performance and enable the algorithm to run in real-time, while sharing resources with other computer vision algorithms. We show application of our algorithm to visual multi-target tracking (MTT) in the presence of parallax and demonstrate its real-time performance on a 640x480 video sequence. Video results are available at https://youtu.be/HhK-p2hXNnU.