Activity understanding and unusual event detection in surveillance videos

Abstract

Computer scientists have made ceaseless efforts to replicate cognitive video understanding abilities of human brains onto autonomous vision systems. As video surveillance cameras become ubiquitous, there is a surge in studies on automated activity understanding and unusual event detection in surveillance videos. Nevertheless, video content analysis in public scenes remained a formidable challenge due to intrinsic difficulties such as severe inter-object occlusion in crowded scene and poor quality of recorded surveillance footage. Moreover, it is nontrivial to achieve robust detection of unusual events, which are rare, ambiguous, and easily confused with noise. This thesis proposes solutions for resolving ambiguous visual observations and overcoming unreliability of conventional activity analysis methods by exploiting multi-camera visual context and human feedback. The thesis first demonstrates the importance of learning visual context for establishing reliable reasoning on observed activity in a camera network. In the proposed approach, a new Cross Canonical Correlation Analysis (xCCA) is formulated to discover and quantify time delayed pairwise correlations of regional activities observed within and across multiple camera views. This thesis shows that learning time delayed pairwise activity correlations offers valuable contextual information for (1) spatial and temporal topology inference of a camera network, (2) robust person re-identification, and (3) accurate activity-based video temporal segmentation. Crucially, in contrast to conventional methods, the proposed approach does not rely on either intra-camera or inter-camera object tracking; it can thus be applied to low-quality surveillance videos featuring severe inter-object occlusions. Second, to detect global unusual event across multiple disjoint cameras, this thesis extends visual context learning from pairwise relationship to global time delayed dependency between regional activities. Specifically, a Time Delayed Probabilistic Graphical Model (TD-PGM) is proposed to model the multi-camera activities and their dependencies. Subtle global unusual events are detected and localised using the model as context-incoherent patterns across multiple camera views. In the model, different nodes represent activities in different decomposed re3 gions from different camera views, and the directed links between nodes encoding time delayed dependencies between activities observed within and across camera views. In order to learn optimised time delayed dependencies in a TD-PGM, a novel two-stage structure learning approach is formulated by combining both constraint-based and scored-searching based structure learning methods. Third, to cope with visual context changes over time, this two-stage structure learning approach is extended to permit tractable incremental update of both TD-PGM parameters and its structure. As opposed to most existing studies that assume static model once learned, the proposed incremental learning allows a model to adapt itself to reflect the changes in the current visual context, such as subtle behaviour drift over time or removal/addition of cameras. Importantly, the incremental structure learning is achieved without either exhaustive search in a large graph structure space or storing all past observations in memory, making the proposed solution memory and time efficient. Forth, an active learning approach is presented to incorporate human feedback for on-line unusual event detection. Contrary to most existing unsupervised methods that perform passive mining for unusual events, the proposed approach automatically requests supervision for critical points to resolve ambiguities of interest, leading to more robust detection of subtle unusual events. The active learning strategy is formulated as a stream-based solution, i.e. it makes decision on-the-fly on whether to request label for each unlabelled sample observed in sequence. It selects adaptively two active learning criteria, namely likelihood criterion and uncertainty criterion to achieve (1) discovery of unknown event classes and (2) refinement of classification boundary. The effectiveness of the proposed approaches is validated using videos captured from busy public scenes such as underground stations and traffic intersections.