Bodytrack(转载)

来源自：parsec benchmark

The bodytrack computer vision application is an Intel RMS workload which tracks a 3D pose of a marker-less(标记点) human body with multiple cameras through an image sequence. This is shown in Figure 1. Bodytrack employs an annealed(退火的) particle(粒子) filter(过滤器) to track the pose using edges and the foreground silhouette(轮廓) as image features, based on a 10 segment 3D kinematic(运动学上的) tree body model. These two image features were chosen because they exhibit a high degree of invariance under a wide range of conditions and because they are easy to extract. An annealed particle filter was employed in order to be able to search high dimensional configuration spaces without having to rely on any assumptions of the tracked body such as the existence of markers or constrained movements。 This benchmark was included due to the increasing significance of computer vision algorithms in areas such as video surveillance(监控), character animation and computer interfaces.

For every frame set of the input videos at time step t, the bodytrack benchmark executes the following steps:

1.The image features of observation are extracted. The features will be used to compute the likelihood of a given pose in the annealed particle filter.

2. Every time step t the filter makes an annealing run through all M annealing layers, starting with layer m=M.

3. Each layer m uses a set of N unweighted particles which are the result of the previous filter update step to begin with. 

Each particle  is an instance of the multi-variate model configuration X which encodes the location and state of the tracked body.