The Reaserch Of Optimal Target-tracking Algorithm Based On Particle Filter Framwork

In those systems such as radar tracking/guidance, intelligent video surveillance androbot vision, many state parameters, including location, velocity and orientation of theinteresting target, are needed to analyze on the kinetic characteristics of the target and torealize further control of it. These state parameters, in most cases, are observed indirectly,and the observation process tends to be interfered by noises. Therefore, methods withfiltering technique should be applied to estimate those state parameters needed on thebasis of the function between observed values and state variables. This type of techniqueis called target-tracking technique. Up to now, different target tracking techniques havebeen applied to aim at different systems, but they are far from being eligible to the realdemand because of their poor robustness and instantaneity.In this paper, we improve Particle Filtering algorithm from the following threeaspects: building a more accurate importance function, resampling strategy and reducingthe computation complexity to resolve the particle digenarete and instantaneity problems.We put forward two target-tracking algorithms based on Particle Filter framework aimingat mixed linear/nonlinear system. Then, the author applies the algorithms in the target-tracking system based on video image. The main research contents are as follows:Firstly, to improve instantaneity of target tracking system, we propose an algorithmcalled Clustering Kalman-prediction PF (CKPF). In this algorithm, Kalman filter is usedto build the importance function by bring in the newest observed value to the importancefunction, to realize a more real condition of sampled particles. Besides, Mean-shiftclustering algorithm is applied to move particles to high posterior probability area. Thus,effectiveness of particles has been improved, evaluated error decreasing and trackingprecision enhancing. Apart from this, as the effective particles have increased, the totalamount of the particles can be reduced, so computation complexity can be lower andinstantaneity is guaranteed.Secondly, in order to improve accuracy of system model and reduce evaluated errorresulting, we establish a combined states model. Depart from color appearance parameter, we introduced location parameter into system model as another system state vector.Furthermore, a combined tracking helps to track the model parameters using the imageparameter: inconsistent movement can be a clue to detect a bad estimation.Thirdly, we split the state into two parts: the location parameters, which are estimatedstochastically, and the color appearance parameters, which are estimated analytically. Bothparts are jointly tracked using Rao–blackwellised technology. In this algorithm, Kalmanfilter is used in prediction and update of color appearance parameters and Particle Filter isused to estimate location parameter of the target. Rao-Blackwellization algorithm reducedthe dimension of mixed linear/nonlinear system and realizing a higher tracking precisionand lower computation complexity. We use KDE method build the target color appearancemodel and updated it with PDA Kalman filter. When ambient brightness and speed oftarget changes or the target is sheltered, the accurs of tracking is still maintaining a highlevel. Hence, robustness of tracing algorithm is improved.Lastly, CKPF is applied in intelligent traffic video surveillance. In this process, thevideo images of intersection captured by high point static camera are taken as researchobjects to realize automatic identification of cutting across solid line road-a traffic offencenow judged mainly by officers. This method can also allow the device to predict theturning of cars at intersections by analyzing the vehicle trajectory, which lays thefoundation for rely tracking of specific vehicles in surveillance cameras network.