基于粒子滤波与神经网络的目标遮挡跟踪

To enhance the reliability of target tracking under occlusion, a tracking strategy based on improved particle filter was proposed. The target histogram model based on kernel function of the hue rank was established to enhance the robustness to the illumination variation. The state equations were established according to the motion information, and the measurement models were established by Radial Basis Function (RBF) neural networks. The Hellinger distance between the template and the target area was used to determine whether the target was under occlusion. When the target was not occluded, the state information of the target was used to update the state equations and train the measurement models. Otherwise, the particle filter was used to fuse the iterative states obtained by state equations and the prediction states obtained by measurement models to get the optimal estimation. Simulation experiments showed that the measurement models based on RBF neural networks could bring some new prediction information which was different to that of the state equations; the strategy could reduce the deviation between the optimal estimation obtained by particle filter and actual states, and enhance the tracking reliability under occlusion.