TY - GEN
T1 - Illumination invariant sequential filtering human tracking
AU - Lu, Yifan
AU - Xu, Dan
AU - Wang, Lei
AU - Hartley, Richard
AU - Li, Hongdong
PY - 2010
Y1 - 2010
N2 - Many tracking problems can be efficiently solved by the Altering technique. Linear filter methods (e.g. Kaiman Filter) have shown their success and optimally in many linear settings with Gaussian noises. However, they expose inefficiency and weakness in the general nonlinear and high dimensional setting (e.g. human tracking). While, the advancement of Sequential Importance Re-sampling with Simulated Annealing has shown it is capable of handling nonlinearity and high dimensionality of human tracking. However, its performance is often affected by lighting variations and noises from silhouette segmentation. The proposed approach incorporates a textured human body template to annealed sequential filtering, and uses the illumination invariant CIELab formula to evaluate the observation likelihood so that influences of lighting changes and noises are minimised. Experiments with the benchmark HumanEval dataset demonstrate encouraging improvements over traditional Sequential Importance Re-sampling and the silhouette based method.
AB - Many tracking problems can be efficiently solved by the Altering technique. Linear filter methods (e.g. Kaiman Filter) have shown their success and optimally in many linear settings with Gaussian noises. However, they expose inefficiency and weakness in the general nonlinear and high dimensional setting (e.g. human tracking). While, the advancement of Sequential Importance Re-sampling with Simulated Annealing has shown it is capable of handling nonlinearity and high dimensionality of human tracking. However, its performance is often affected by lighting variations and noises from silhouette segmentation. The proposed approach incorporates a textured human body template to annealed sequential filtering, and uses the illumination invariant CIELab formula to evaluate the observation likelihood so that influences of lighting changes and noises are minimised. Experiments with the benchmark HumanEval dataset demonstrate encouraging improvements over traditional Sequential Importance Re-sampling and the silhouette based method.
UR - http://www.scopus.com/inward/record.url?scp=78149329165&partnerID=8YFLogxK
U2 - 10.1109/ICMLC.2010.5580491
DO - 10.1109/ICMLC.2010.5580491
M3 - Conference contribution
SN - 9781424465262
T3 - 2010 International Conference on Machine Learning and Cybernetics, ICMLC 2010
SP - 2133
EP - 2138
BT - 2010 International Conference on Machine Learning and Cybernetics, ICMLC 2010
T2 - 2010 International Conference on Machine Learning and Cybernetics, ICMLC 2010
Y2 - 11 July 2010 through 14 July 2010
ER -