Research Of Multi-Target Tracking Algorithms

Visual target tracking is an important topic in computer vision with many applica-tions,which has attracted wide attention of the researchers at home and abroad. Mul-ti-target tracking is a fundamental task for higher level automated video content analysis,where the number of targets is not certain and the interaction exists between different targets, and it is a difficult problem in visual tracking. Although many new algorithms have been proposed for multi-target tracking recent years, it remains a big challenge to get stable and accuracy tracking results. So, this paper does an in-depth research in mul-ti-target visual tracking, which includes two main aspects: multi-target filtering and asso-ciation, and some practical and effective methods are proposed for solving the difficulty in tracking.Firstly, this paper researches in Probability Hypothesis Density (PHD) filter based on random finite sets (RFS), which can estimate the states of multi-targets. For a start, this approach gets the observations by video object detectors, and models the states and observations as RFS, and then computes the PHD in every frame iteratively, which indi-cates the number and estimated states of multi-targets. Tacking this idea a few steps fur-ther, this paper proposes a Kalman particle PHD filter for visual tracking. We perform Kalman filter in the prediction stage of the PHD particle filter and obtain a new proposal distribution, which is close to the posterior of targets and minimizes the covariance of importance sampling weights. For the update, we adopt a new observation model, which considers both the object dynamic states and the object appearances, and derives a more precise likelihood.For the generation of multi-target tracklets, this paper proposes a 2-level association algorithm based on online boosting. After the detection of multi-target in every frame, we collect the training samples from initialized tracklets and create Online Learned Dis-criminative Appearance Models (OLDAMs), and then train classifiers with boosting al-gorithm, which is applied to associate the targets. For solving the problem that the track- lets obtained by classifier are too short and trivial to get a continuous tracking, we per-form a second association on the short tracklets. We formulate it as an optimal matching problem of bipartite graph and solve it with Kuhn-Munkras (KM) algorithm, and then we get an optimal association set, which results in a stable and accurate tracking for multi-target.Furthermore, we simulate our algorithms on real-world scenarios, the performance of which is compared with other methods. Experimental results demonstrate that our proposed methods have advantages in multi-target visual tracking, and often outperform the state of the art on the popular datasets.