Title: "Visual Vehicle Tracking Using Direction-based Particle Filter with Online Threshold Adaptation"
         

DOI: 10.15224/978-1-63248-113-9-09
Page(s): 41 - 45
Authors: BURAK MERDENYAN, MUSTAFA EREN YILDIRIM, YUCEL BATU SALMAN

Abstract

In various applications of computer vision, nonlinearity and non-Gaussianity must be considered to have an accurate and realistic modeling. In this sense, particle filter is preferred for tracking. In this paper, a modified particle filter which takes the direction of the vehicle into account and distributes the particles according an automatic thresholding scheme is used. After obtaining the vehicle direction, particles are weighted according to their angular similarities to the vehicle. Particles having angular distance greater than a threshold are eliminated. Thus, the remained particles that are moving in same or similar direction with vehicle, increase their own probability of likelihood. This threshold is decided online while the algorithm runs. Depending on the number of particles used, this scheme increases or decreases the threshold. This scheme prevents the system from failures due to insufficient number of particles and also from using excessive number of particles. Proposed algorithm is compared with condensation algorithm and Camshift algorithm. According to the results, proposed algorithm surpasses the other algorithms in terms of tracking period and precision.