“High performance template tracking using fixed models” by Cabido, Montemayor, Pantrigo and Payne

Visual tracking consists in the estimation and prediction of a target state. One of the most popular approaches for visual tracking is the particle filters (PF) framework. In a PF framework the state of a target (location, size, angle, etc.) at time t is approximated by a set of estimates (particles), where each one is composed of a state of interest and a weight factor, pi=(xi,πi). In our 2D visual tracking problem the state is x=(x, y, θ, s) (position, rotation and scale). Particle filters do not deal well with high dimensional estimation problems. In these cases, they need from an exponential increase in the number of particles for the same tracking accuracy and, there- fore, much more computational cost. In this work we propose the graphics processing unit computational power to overcome a 2D tracking problem in a PF framework for more than real time performance (30 fps). We present a complete GPU rotation-translation-scale visual tracking solution that uses a previously acquired template for the evaluation stage of a PF. The key point is that the GPU offers hardware acceleration for interpolation and texture sampling that can be exploited by the application. It delivers much higher processing rates than interactive video playback enabling more robustness for the tracking problem. This GPU solution would be very promising when using future optimization strategies, such as adaptive appearance models for visual tracking [Ross et al. 2007].

1. Montemayor, A. S., Pantrigo, J. J., Cabido, R., Payne, B., Sánchez, A., and Fernández, F. 2006. Improving GPU Particle Filter with Shader Model 3.0 for Visual Tracking. In Proc. of ACM SIGGRAPH 2006-Research Posters.
2. Ross, D., Lim, J., Lin, R., and Yang, M. 2007. Incremental learning for robust visual tracking. International Journal of Computer Vision. Special Issue: Learning for Vision.