Abstract
We formulate human motion tracking as a high-dimensional constrained optimization problem. A novel generative method is proposed for human motion tracking in the framework of evolutionary computation. The main contribution is that we introduce immune genetic algorithm (IGA) for pose optimization in latent space of human motion. Firstly, we perform human motion analysis in the learnt latent space of human motion. As the latent space is low dimensional and contents the prior knowledge of human motion, it makes pose analysis more efficient and accurate. Then, in the search strategy, we apply IGA for pose optimization. Compared with genetic algorithm and other evolutionary methods, its main advantage is the ability to use the prior knowledge of human motion. We design an IGA-based method to estimate human pose from static images for initialization of motion tracking. And we propose a sequential IGA (S-IGA) algorithm for motion tracking by incorporating the temporal continuity information into the traditional IGA. Experimental results on different videos of different motion types show that our IGA-based pose estimation method can be used for initialization of motion tracking. The S-IGA-based motion tracking method can achieve accurate and stable tracking of 3D human motion.
Highlights
Tracking articulated 3D human motion from video is an important problem in computer vision which has many potential applications, such as virtual character animation, human computer interface, intelligent visual surveillance, and biometrics
We propose an immune genetic algorithm (IGA)-based method for pose estimation from monocular images
Experimental results demonstrate that our sequential IGA (S-IGA)-based tracking method can achieve accurate and stable tracking of 3D human motion
Summary
Tracking articulated 3D human motion from video is an important problem in computer vision which has many potential applications, such as virtual character animation, human computer interface, intelligent visual surveillance, and biometrics. Three-dim pose Static images Body model Image sequence space [7, 8] Motivated by this observation, we use ISOMAP, a nonlinear dimensionality reduction method, to learn the lowdimensional latent space of pose state, by which the aim of both reducing dimensionality and extracting the prior knowledge of human motion are achieved simultaneously. Search strategy, in general how to track in the low-dimensional latent space, is another important problem. We propose a novel generative approach in the framework of evolutionary computation, by which we try to widen the bottlenecks mentioned above with effective search strategy embedded in the extracted state subspace. As the latent space is low dimensional and contents the prior knowledge of human motion, it makes pose analysis more efficient and accurate.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
