|
Automatic assessment of situations with modern security and surveillance systems requires sophisticated discrimination capabilities. Therefore, action recognition, e.g. in terms of person-person or person-object interactions, is an essential core component of any surveillance system. A subclass of recent action recognition approaches are based on space time volumes, which are generated from trajectories of multiple anatomical landmarks like hands or shoulders. A general prerequisite of these methods is the robust estimation of the body pose, i.e. a simplified body model consisting of several anatomical landmarks. In this paper we address the problem of estimating 3D poses from monocular person image sequences. The first stage of our algorithm is the localization of body parts in the 2D image. For this, a part based object detection method is used, which in previous work has been shown to provide a sufficient basis for person detection and landmark estimation in a single step. The output of this processing step is a probability distribution for each landmark and image indicating possible locations of this landmark in image coordinates. The second stage of our algorithm searches for 3D pose estimates that best t to the 15 landmark probability distributions. For resolving ambiguities introduced by uncertainty in the locations of the landmarks, we perform an optimization within a Particle Swarm Optimization (PSO) framework, where each pose hypothesis is represented by a particle. Since the search in the high-dimensional 3D pose search space needs further guidance to deal with the inherently restricted 2D input information, we propose a new compact representation of motion sequences provided by motion capture databases. Poses of a motion sequence are embedded in a low-dimensional manifold. We represent each motion sequence by a compact representation referred to as pose splines using a small number of supporting point poses. The PSO algorithm can be extended to perform the optimization process directly on pose splines. Results of the proposed method are shown on the UMPM benchmark.
|
