| Tracking multiple humans in real scenes is an important problem in computer vision due to its importance for many applications, such as surveillance, robotics, and human-computer interactions. Association based tracking often achieves better performances than other approaches in crowded scenes. Based on this framework, I propose offline and online learning algorithms to automatically find potential useful appearance and motion patterns, and utilize them to deal with difficulties in the association framework and to produce much better tracking results.;In association based framework, an offline learned detector is first applied in each video frame to produce detection responses, which are further associated into tracklets, i.e., track fragments, in multiple steps. Measurement of affinities between tracklets is the key issue that determines the performance. In the first part of my thesis, I propose an online learning algorithm which automatically find three important cues from a static scene to improve tracking performance: non-linear motion patterns, potential entry/exit points, and co-moving groups.;Association based tracking methods are often based on the assumption that affinities between tracklet pairs are independent of each other. However, this is not always true in real cases. In order to relax the independent assumption, we introduce an offline learned Conditional Random Field (CRF) model to estimate both affinities between tracklets and dependencies among them. Finding best associations between tracklets is transformed into an energy minimization problem, and energies of unary and pairwise terms in the CRF model are offline learned from pre-labeled ground truth data by a RankBoost algorithm. Then I further extended the approach into an online version. Positive and negative pairs are online collected according to temporal constraints; the learned appearance models better distinguish close but visually similar targets and the learned motion models considered relative distances between targets to alleviate camera motion and non-linear path effects.;As detection performance limits the performances of traditional association based tracking approaches, I further propose an online learned discriminative part-based appearance models which incorporates category free tracking techniques into association based tracking. In this work, occlusions among targets are explicitly considered to produce more robust appearance models. A category free tracking method is adopted to track a target without detection responses while distinguishing different targets and the background.;I designed comprehensive experiments to evaluate all my algorithms and important modules. The performances show effectiveness of my approaches on different data sets, with different human densities, illuminations, camera motions, and etc. |
