Modelling and estimation of vehicle tracking using and improved particle filter

• Main Author: Khong, Wei Leong
• Format: Thesis
• Language: English; English
• Published: 2013
• Subjects: TK5101-6720 Telecommunication Including telegraphy, telephone, radio, radar, television
• Online Access: https://eprints.ums.edu.my/id/eprint/41460/1/24%20PAGES.pdf; https://eprints.ums.edu.my/id/eprint/41460/2/FULLTEXT.pdf; https://eprints.ums.edu.my/id/eprint/41460/
• Institution: Universiti Malaysia Sabah

This research focuses on reducing the particle size in the resampling stage of the particle filter approach by tracking a single vehicle with overlapping situation. Particle filter is competent to robustly tracking the vehicle under various situations. The vehicle can be tracked by estimating the position of the target vehicle with a set of distributed random particles with associated weight. Since the estimated position is computed based on the mean value of the hypotheses, the accuracy and efficiency of the particle filter are greatly affected by the particle size. Besides, the placement of the particles also plays an important role in producing accurate tracking results. In practice, the conventional particle filter is facing the particle degeneracy problem after a few iteration of the estimation process. Although the resampling stage in particle filter can overcome the particle degeneracy problem, the large number of particles required to resample is uncertain due to the encountered occlusion situation. Hence, a genetic algorithm based resampling technique will be embedded into the particle filter algorithm to reduce the amount of the resampling particles and subsequently reduce the particle size. Based on the nature of the genetic algorithm, a better estimation of position of the target vehicle can be obtained by recombining the information between the particles. With the improvement of the particle placement, the number of particles used in the resampling stage can be reduced and hence decrease the iteration of the resampling process. Results show that the particle filter with genetic algorithm resampling has successfully reduced 45.5 % of the particles in the resampling stage before the target vehicle is fully occluded by the obstacle vehicle. Subsequently, the developed algorithm has also reduced 50.2 % of the resampling particles when the target vehicle reappears but still partially occluded from the occlusion as compared to the fundamental resampling approach.