Comparison between GMM and KDE data fusion methods for particle filtering: Application to pedestrian detection from laser and video measurements

Abstract

In urban environment, pedestrian detection is a challenging task in automotive research, which often suffers from the lack of reliability due to the occurrences of spurious detections. In order to answer multitarget multisensor tracking problem and more specifically pedestrian tracking, we propose to use an algorithm based on a stochastic recursive Bayesian framework also called particle filter. We aim to solve the problem of consistent Bayesian Decentralized Data Fusion (BDDF) with particle filter using two different statistics approaches in order to better represent the particle set and maintains an accurate summary of the particles. We propose a comparison between a Kernel Density Estimation (KDE) based on non-parametric estimation and a Gaussian Mixture Model (GMM) based on parametric estimation. This approach allows to cope with non-linear models and multi-modalities induced by occlusions and clutters. These two algorithms differ in the representation of particle set during data fusion. Simulation results as well as the results of the experiments conducted on real data demonstrate the relevance of these approaches.