Trajectory based video analysis in multi-camera setups

This thesis presents an automated framework for activity analysis in multi-camera setups. We start with the calibration of cameras particularly without overlapping views. An algorithm is presented that exploits trajectory observations in each view and works iteratively on camera pairs. First outliers are identified and removed from observations of each camera. Next, spatio-temporal information derived from the available trajectory is used to estimate unobserved trajectory segments in areas uncovered by the cameras. The unobserved trajectory estimates are used to estimate the relative position of each camera pair, whereas the exit-entrance direction of each object is used to estimate their relative orientation. The process continues and iteratively approximates the configuration of all cameras with respect to each other. Finally, we refi ne the initial configuration estimates with bundle adjustment, based on the observed and estimated trajectory segments. For cameras with overlapping views, state-of-the-art homography based approaches are used for calibration. Next we establish object correspondence across multiple views. Our algorithm consists of three steps, namely association, fusion and linkage. For association, local trajectory pairs corresponding to the same physical object are estimated using multiple spatio-temporal features on a common ground plane. To disambiguate spurious associations, we employ a hybrid approach that utilises the matching results on the image plane and ground plane. The trajectory segments after association are fused by adaptive averaging. Trajectory linkage then integrates segments and generates a single trajectory of an object across the entire observed area. Finally, for activities analysis clustering is applied on complete trajectories. Our clustering algorithm is based on four main steps, namely the extraction of a set of representative trajectory features, non-parametric clustering, cluster merging and information fusion for the identification of normal and rare object motion patterns. First we transform the trajectories into a set of feature spaces on which Meanshift identi es the modes and the corresponding clusters. Furthermore, a merging procedure is devised to re fine these results by combining similar adjacent clusters. The fi nal common patterns are estimated by fusing the clustering results across all feature spaces. Clusters corresponding to reoccurring trajectories are considered as normal, whereas sparse trajectories are associated to abnormal and rare events. The performance of the proposed framework is evaluated on standard data-sets and compared with state-of-the-art techniques. Experimental results show that the proposed framework outperforms state-of-the-art algorithms both in terms of accuracy and robustness.EThOS - Electronic Theses Online ServiceGBUnited Kingdo