Citation Context

...], ut is a measure of uniformity for P(ot|·) over this range and the numerator in (3c) is forced to the range [−180 ◦ ..180 ◦ ]. ut is in fact the survival diagnostic (from particle filter resampling =-=[5, 6]-=-). Note that �ωt ∝1−ut, which is 0 for the uniform distribution and 1 for the impulse distribution and is exactly the weighting behaviour desired. Equation (3c) then adjusts the weights to take into a...

Citation Context

... length of each limb is fixed. Each body part is modelled with a cylinder whose sides are projected onto the 2D image and then joined with lines to produce the cardboard look for efficient projection =-=[12]-=-. The model is fairly loose-fitting so that any tracker based on it should generalise well to different people. Scale 1 df 3 df 1 df 1 df 3 df 2 df 3 df 3 df 3 df 1 df 1 df 1 df z y x Orientation/Pitc...

Citation Context

...d from P(xk t |q (i) t ) (i.e. xk t ⊥ xk t−1 if ek t=1). This uses context-specific independence on xk [2], or alternatively could be viewed as a mixture of Monte Carlo kernels for inference purposes =-=[1]-=-. Formally: e k t = sample x k(i) t ∼ ( 1 if ∀j, P(x k t =j, q (i) t |x k(i) t−1 ,q (i) t−1 )=0 0 otherwise ( A k nij � P(x k t |q (i) t ,x k(i) t−1 ,e k t =0) Λnj � P(x k t |q (i) t ,e k t =1) t |ek ...

Citation Context

... is allowed. If a sub-state xk t−1 cannot transition into qt, ek t is set to 1 and xk t is sampled from P(xk t |q (i) t ) (i.e. xk t ⊥ xk t−1 if ek t=1). This uses context-specific independence on xk =-=[2]-=-, or alternatively could be viewed as a mixture of Monte Carlo kernels for inference purposes [1]. Formally: e k t = sample x k(i) t ∼ ( 1 if ∀j, P(x k t =j, q (i) t |x k(i) t−1 ,q (i) t−1 )=0 0 other...

Citation Context

...resolution and frame rate. 1. Introduction A variety of approaches to the problem of markerless 3D full-body human motion capture have been proposed in the literature. Lee et al. [7] and Mikić et al. =-=[8]-=- both constrain the possible posture configurations by analytically finding the hands, face and/or torso. Lee then transitions a particle filter under these constraints while Mikić ‘grows’ the bodypar...

Citation Context

...e space, each action is broken down into a two-level hierarchy of phases (sub-actions) and motion within each phase. The hierarchy is tractably modelled with a hierarchical hidden Markov model (HHMM) =-=[3]-=- by factoring the states of the lower level (which model the actual pose). Each action is then modelled by a different instance of this factored-state HHMM (FS-HHMM), and the most likely model for a g...

Citation Context

...], ut is a measure of uniformity for P(ot|·) over this range and the numerator in (3c) is forced to the range [−180 ◦ ..180 ◦ ]. ut is in fact the survival diagnostic (from particle filter resampling =-=[5, 6]-=-). Note that �ωt ∝1−ut, which is 0 for the uniform distribution and 1 for the impulse distribution and is exactly the weighting behaviour desired. Equation (3c) then adjusts the weights to take into a...

Citation Context

...cking with a reduced resolution and frame rate. 1. Introduction A variety of approaches to the problem of markerless 3D full-body human motion capture have been proposed in the literature. Lee et al. =-=[7]-=- and Mikić et al. [8] both constrain the possible posture configurations by analytically finding the hands, face and/or torso. Lee then transitions a particle filter under these constraints while Miki...

Citation Context

...ng to gradually focus the search effort on promising areas. The algorithm is effective but tends to converge on only one mode, discarding the rest of the posture distribution. Sminchisescu and Jepson =-=[13]-=- explicitly maintain multi-modality by using a combination of kinematic jumps, sampling and variational methods to track and smooth multiple plausible posture trajectories. Their system is able to rec...

Citation Context

...most likely model for a given sequence provides the action label and posture sequence. The basic approach of the FS-HHMM to tracking is similar to Zhao and Nevatia’s ‘tracking-as-recognition’ concept =-=[14]-=-. However, they combine tracking and recognition by matching optic flow against labelled motion templates and filter (track) with an HMM to produce a maximumlikelihood sequence of motion. Also, templa...

Citation Context

...any other posture trajectories at the same time) at the expense of a complex, multi-layered algorithm structure and an implicit reliance on a close-fitting body model. Recent work by Caillette et al. =-=[4]-=- learns Gaussian clusters of sub-motions and trains a variable-length Markov model (VLMM) based on these clusters to direct the local posture search towards better areas of the distribution. They achi...

Citation Context

...nd empirically set based on how fast a person is expected to move with respect to the video frame rate. The final two parameters are extracted by bootstrapping the FS-HHMM from a bounding box tracker =-=[11]-=-. Bootstrapping is based on the assumption that a person will walk upright into the room facing forward, thus implying scale and orientation. The system waits to bootstrap the FS-HHMM until the box-tr...

Citation Context

...e this paper aims to handle errors, it employs the latter approach: f(yt|xt)= 1 λ exp˘λ · Dist(yt,Proj(xt)) ¯ (1) where Dist(·) is a modified version of Deutscher’s [5] cost function (as described in =-=[10]-=-) for the distance between yt and xt’s projection and λ controls how sharply the distribution drops off with distance. For the FS-HHMM λ is fixed with λ=8 (chosen empirically). For the APF λ is varied...

Citation Context

...· βt(h) h A k=1 k nij “X ” Y24 αt(n)= αt−1(m) · Cmn A k nij, α1(n)=φn m βt(m)= X“ Y24 βt+1(n) · Cmn n k=1 k=1 A k nij (6a) (6b) 24Y π k ni k=1 (6c) ” , βT(m)�1 (6d) where i � x k t−1, j � x k t . See =-=[9]-=- for details of the derivation. Inference and Auto-Initialisation As well as the learned model parameters, the FS-HHMM requires three additional parameters to perform inference: (i) variance of P(ot|o...