Target Tracking By A New Class Of Cost-Reference Particle Filters

Standard particle filters have shown excellent performance in many challenging scenarios of target tracking, and therefore they often are the method of choice. In cases when there is no knowledge about the noise distributions in the studied system, one cannot use these methods or will use them with assumptions that in general may lead to very poor results. An alternative to standard particle filters are the cost-reference particle filters. They are also based on the principle of exploring the state-space by drawing particles in that space but they do not require probabilistic information about the system. As with all particle-based filters, an important step in the implementation of cost-reference particle filters is the generation of new particles. In this paper we propose a new class of cost-reference particle filters which uses the extended Kalman filter for drawing of candidate particles. We demonstrate the performance of these filters on target tracking problems. We compare the new filter with traditional ones by simulated experiments.