A Quantitative Evaluation of Video-based 3D Person Tracking

Abstract

The Bayesian estimation of 3D human motion from video sequences is quantitatively evaluated using synchronized, multi-camera, calibrated video and 3D ground

Tracking a walking person using activity-guided annealed particle filtering

“…Balan et al [3] made a quantitative study of APF performance, finding an average lower bound on absolute joint location error of 41mm for a walking subject using 1000 particles and 3 cameras. However, tracking was found to fail for the 2-camera and monocular cases after approximately 1 second, despite equivalent particle numbers.…”

“…In this work, we use the synchronised video and MoCap dataset described in [3] in order to draw direct comparisons with standard APF performance. The dataset contains multi-camera synchronised video sequences of a walking subject taken at 60fps.…”

“…Research into the markerless tracking of human motion has recently benefitted from the introduction of common data sets that include ground truth motion capture (MoCap) data [3,10]. These have allowed for the quantitative evaluation and cross-comparison of tracking approaches.…”

“…These have allowed for the quantitative evaluation and cross-comparison of tracking approaches. Annealed Particle Filtering (APF) [4] and Sampling Importance Resampling (SIR) [1] have been shown to recover pose from multiple cameras using silhouette and edge cues [3]. However, both approaches fail when limited to observations from only 1 or 2 cameras.…”

“…We test the HMM-guided APF (HMM-APF) approach on the walking dataset presented in [3] and evaluate against ground truth MoCap data. Known and unknown subject tracking is demonstrated from 2 cameras using 200 particles.…”