Particle filter based visual tracking with multi-cue adaptive fusion

To improve the robustness of visual tracking in complex environments such as: cluttered backgrounds, partial occlusions, similar distraction and pose variations, a novel tracking method based on adaptive fusion and particle filter is proposed in this paper. In this method, the image color and shape cues are adaptively fused to represent the target observation; fuzzy logic is applied to dynamically adjust each cue weight according to its associated reliability in the past frame; particle filter is adopted to deal with non-linear and non-Gaussian problems in visual tracking. The method is demonstrated to be robust to illumination changes, pose variations, partial occlusions, cluttered backgrounds and camera motion for a test image sequence.     

Track before detect for point targets with particle filter in infrared image sequences

The problem of detecting and tracking point targets in a sequence of infrared images with very low signalto-noise ratio (SNR) is investigated in this paper. A track before detect algorithm for infrared (IR) point target is developed based on particle filter. The particle filter is used to estimate the state of the target in track stage. The unnormalized weights of the output of the filter are used to approximately construct the likelihood ratio for hypothesis test in detection stage. Experiment results with the real image sequences that SNR is about 2.0 show that the proposed algorithm can successfully detect and track point target.     

Modified maximum likelihood registration based on information fusion

The bias estimation of passive sensors is considered based on information fusion in multi-platform multi- sensor tracking system.The unobservable problem of bearing-only tracking in blind spot is analyzed.A modified maximum likelihood method,which uses the redundant information of multi-sensor system to calculate the target position,is investigated to estimate the biases.Monte Carlo simulation results show that the modified method eliminates the effect of unobservable problem in the blind spot and can estimate the biases more rapidly and accurately than maximum likelihood method.It is statistically efficient since the standard deviation of bias estimation errors meets the theoretical lower bounds.