Adaptive visual tracking by decision level fusion of features in a particle filtering frame work

In this paper we propose a new method for multi-feature object tracking in a particle filter framework. Each particle indicates one hypothesis of tracked object. In common method of feature combination, each particle measures all features. Due to limited computational power, particle filter is forced to run with lower number of particles which results in weak approximation of posterior distribution of target state. In our method, each hypostasis is evaluated by only one feature, so the number of particles can be increased. In our method the percentage of the particles which participate in the measurement of a specific feature is directly related to the reliability of that feature. Measuring more reliable features and having better spatially distributed samples from the scene are two main advantages of our method. Experimental results over a set of real-world sequences demonstrate better performance of our method compared to common method of feature combination.

Keywords:

feature combination , feature reliability , object tracking , particle filter , video sequences