3d Human Motion Tracking Based On Annealed Particle Filtering In Monocular Video

Three-dimensional human motion tracking is an important research domain in machine vision field in recent years. It has a wide range of applications, such as human-computer interaction, monitoring of intelligent video and animation etc. At present, many research topics about three-dimensional human motion tracking are based on multiple video sequences, however, monocular video sequences couldn't reach a satisfied accuracy in resuming human body three-dimensional attitude data because of monocular video lacking of detailed information and self-occlusion between the body parts. Concerning on this issue, we choosed particle filtering method based on the Monte Carlo methods, combined with prior knowledge of human movement to realize human body three-dimensional tracking.We add simulated annealing to reduce the number of particles. In the tracking process, we took advantage of weighted particles collection to approach the posterior probability distribution, and adjusted the location of particles from time to time, let them close to the peak of likelihood function gradually , in order to make the weighted particles to have a better reflection of distribution in each moment.We used two kinds of image features edge and silhouette to construct likelihood function, making the quantification-comparability estimation between model projections and images more accurate to achieve the purpose of determining the importance of particles reasonablely, which makes robustness of tracking results more finer.Integrate the constrained human motion into the entire track framework, including the perspective scope restriction of the skeleton movement constrained and non-penetrating restriction of limbs in order to restrict the feasible area of attitude data for preventing the emergence of illegal attitude data and improving the accuracy of three-dimensional human tracking. In addition, add a simple linear prediction model into the process of tracking, and verified the influence of two linear movement prediction models to tracking effect.Experiments show that the method mentioned above are effective in 100 frames for walking motion tracking with rotary under the complex background , while it's difficult for the method to track 3D human motion in long time sequence which need adding more complex motion prediction models and human movement restrictions to be superintended.