Improved particle filter in sensor fusion for tracking random moving object

Abstract

Non-linear and non-Gaussian estimation is a challenging problem in multi-sensor fusion. To handle this, particle filter is used to estimate the system state based on information from camera and sonar sensor. The state variables such as position, velocity and acceleration of a random moving object change very quickly and are hard to track. This leads to serious sample impoverishment in particle tracking algorithm. In this paper, a new resampling algorithm is presented. Random samples are drawn from the neighborhoods of previous samples with high weights and the effect of sample impoverishment is reduced. The state space model is augmented with acceleration variables to describe the random movement more accurately.