Fast nonparametric belief propagation for real-time stereo articulated body tracking

Combination of Annealing Particle Filter and Belief Propagation for 3D Upper Body Tracking

“…Human body is then represented as a graphical model where individual limbs are characterized by nodes and relationships between body parts are represented by edges connecting nodes and encoded by conditional probability distributions. This graphical model allows tracking each subpart individually, and then adding constraints between adjacent limbs thanks, for example, to Belief Propagation (BP) inference algorithm [32,33,34,21,35]. In this way, the high dimensionality problem is expressed as a set of lower dimension, and thus the complexity of the search task is linear, rather than exponential, according to the number of body parts.…”

“…This graphical model allows to track each subpart individually, and then to add constraints between adjacent limbs. By doing so, it was possible for us to add the Belief Propagation (BP) inference algorithm [21,[35][36][37][38][39]. In this way, the initial high dimensionality problem is expressed as several problems of lower dimension, and thus the complexity of the search task is linear rather than exponential according to the number of body parts.…”

“…• the interaction potentials (or link factors) Then the joint probability at time k can be written as a product of independent factors [35]:…”

Swapping-Based Partitioned Sampling for Better Complex Density Estimation: Application to Articulated Object Tracking

“…It combines the PF framework with the well-known Loopy Belief Propagation algorithm [14] for speeding-up computations (but at the expense of approximations). It has been successfully applied on many problems of high dimensions [16,2,7] Another popular approach is the Rao-Blackwellized Particle Filter for DBN (RBPF) [5]. By using a natural decomposition of the conditional probability, RBPF decomposes the state space into two parts that fulfill the following condition: the conditional distribution of the second part given the first part can be estimated using classical techniques such as Kalman filter.…”

Probabilistic 3D Tracking: Rollator Users' Leg Pose from Coronal Images

“…With regard to multiple cameras, 3D limb tracking has also been widely addressed and reliably implemented (e.g. [1]). Hardware portability, limited power supply and space constraints prevent us from installing a stereo vision system on the walker.…”