Research Of Video Target Tracking Based On Particle Filtering Framework

Video target tracking is one of the most important and popular issues in the field ofcomputer vision, which has significant applications in military operation, trafficmonitoring, medical diagnosis and so on. In this thesis, the Particle Filter will play therole of tracking framework. The research will be further carried out from the respects ofboth tracking framework itself and the target feature. The former constructs the generalarchitecture of the tracking algorithm, while the later makes the algorithm workable formany different tracking environments.In this thesis, we first introduce the theory of Particle Filtering and how it works asa tracking framework. Then a classical mode seeking algorithm Mean Shift, which willbe applied as an optimizing method into the framework and edge feature matchingprocess, is simply described. After that, the features of color, edge, and contour will beintroduced separately, including the details of feature abstracting and matching underthe Particle Filtering framework. At last, we will assemble both the features andoptimizing details into the framework, and obtain the complete tracking algorithm. Andthe performance of such algorithm is tested during experiments. Now we present whatwe did and what we gain during the whole research.In the case of tracking framework, the Particle Filter uses the motion of target as thesystem state transfer model. Traditionally, the target motion is under an assumption ofuniform speed which is computed as the mean speed of a certain number of latestframes. Such motion will probably lead to a lagged tracking. In this thesis, it isimproved by calculating the weighted average speed rather than the pure average speed.Moreover, a two-order approximated motion will be applied into the system statetransfer model so that the forecasted result can fit the ground truth much better.Considering the problem of heavy computation in Particle Filter, we use Mean Shiftto cluster the particles, and move such process ahead of re-sampling, so that both thepositions and weights of particles can be clustered, which not only raises the expressingability of the particle set and thus reduce the number of all particles, but also maintainsmore weight information during the clustering than the traditional method does.In the case of target feature, we pick three classical features and fuse them to obtainhighly adaptable and robust tracking. For the color feature, we apply an improvedlikelihood measurement which is much more suitable for computer simulation and, to a certain extent, can raise the matching accuracy and save the time.For the edge feature, we proposed a novel matching method to overcome theproblem of poor effectiveness of traditional way. Such method uses a series ofgray-level morphological operations to extract and smooth the edge feature, and appliesMean Shift to adjust the edge information of each model to the mode where the edgedistribution gets to a local peek, after that the matching process is finally implemented.The likelihoods obtained by using such method can reflect the actual distribution ofcandidate models. Moreover, for different tracking backgrounds, we also proposecorresponding strategies to keep a satisfying tracking performance.Furthermore, we add contour as the third feature to the algorithm, so that thetracking performance can get further improved. At last, we fuse the three featuretracking results to obtain the final target position. Experimental results show that, ouralgorithm can combine the merit of each feature and realize effective tracking in a lot ofdifferent environments. Meanwhile, based on the forecasting character of ParticleFiltering framework and the feature fusing strategy, the tracking algorithm can stillwork well when the target is occluded temporarily or one feature loses its distinctness.