Neural-bayesian Filtering Based on Monte Carlo Resampling for Visual Robust Tracking

The visual robust tracking is an acid test for existing methods since the target with large-dynamic-change scenarios. Specifically, this paper presents neural aided Bayesian filtering scheme which is based on Monte Carlo resampling techniques associated with lower particles hypothesis to addresses the computational intensity that is intrinsic to all particle filter (PF) approaches, including those that have been modified to overcome the degeneracy of particles and improve the diversity of particle samples. Performance and tracking quality results for severe-dynamic target tracking experiments demonstrate that the Bayesian filtering deteriorated error caused by constrain the number of particles required which was compensated by a RBF neural network, with high accuracy and intensive tracking performance for unconstrained abrupt motion only require lower particles compare with SIR Bayesian filtering, meanwhile, the proposed method is also with strong robustness for different number of particles.