Variational Bayesian PHD Filter with Deep Learning Network Updating for Multiple Human Tracking

Abstract

We propose a robust particle probability hypothesis density (PHD) filter where the variational Bayesian method is applied in joint recursive prediction of the state and the time varying measurement noise parameters. The proposed particle PHD filter is based on forming variational approximation to the joint distribution of states and noise parameters at each frame separately; the state is estimated with a particle PHD filter and the measurement noise variances used in the update step are estimated with a fixed point iteration approach. A deep belief network (DBN) is used in the update step to mitigate the effect of measurement noise on the calculation of particle weights in each frame. The deep learning network is trained based on both colour and oriented gradient histogram (HOG) features and then used to mitigate the measurement noise from the particle selection step, thereby improving the tracking performance. Simulation results using sequences from the CAVIAR dataset show the improvements of the proposed DBN aided variational Bayesian particle PHD filter over the traditional particle PHD filter.