adaptive sparse representations for video anomaly detection

Open Access
Author:
Mo, Xuan
Graduate Program:
Electrical Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
June 16, 2014
Committee Members:
  • Vishal Monga, Dissertation Advisor
  • Vishal Monga, Committee Chair
  • William Evan Higgins, Committee Member
  • Kenneth Jenkins, Committee Member
  • Jesse Louis Barlow, Committee Member
  • Raja Bala, Special Member
Keywords:
  • video anomaly detection
  • sparsity model
  • kernel function
  • multi-object
  • low rank sparsity prior
  • outlier rejection
Abstract:
Video surveillance systems are widely used in the transportation domain to identify unusual patterns such as traffic violations, accidents, unsafe driver behavior, street crime, and other suspicious activities. As the volume of video data increases, most video analysis requires significant human supervision. However, since human supervision is not scalable, a software-aided real-time video surveillance system is desirable. Video anomaly detection problem has been formulated as a well-known likelihood ratio test (LRT) problems, under the idealized assumption of the knowledge of the distributions of both normal events and anomalous events. Often, ample training corresponding to anomalous events is assumed to be available (supervised setting). In many real-world problems, however, both normal and anomalous distributions are generally unknown and difficult to estimate even when the training data is available. The key open challenges in video anomaly detection comprise: 1.) presence of noise in surveillance videos and occlusions in transportation videos; 2.) detection of anomalies involving multiple objects; 3.) lack of training samples representing anomalous events; 4.) representation of event using multiple features (known formally as video event encoding); 5.) identification of anomalies in unstructured scenario where preparation of a dictionary clearly separated into class-specific sub-dictionaries is impossible. Recently sparse reconstruction techniques have been used for image classification, and shown to provide excellent robustness to occlusion. This progress has also been leveraged for sparsity-based video-anomaly detection where test events are expressed as sparse linear combinations of example events from a given (normal or anomalous) class. This dissertation explores novel and adaptive sparse representations for addressing open challenges in video anomaly detection. First, we develop a new joint sparsity model for anomaly detection that enables the detection of joint anomalies involving multiple objects and then we propose outlier rejection measure for unsupervised video anomaly detection. Second, we introduce non-linearity into the linear sparsity model and dictionary design and optimization technique to enable superior class separability and dictionary compactness. Third, we propose to extend sparsity models based on single feature representations to more sophisticated sparse models based on multiple feature representations. Finally, we introduce low rank sparsity prior to our sparsity model which perfectly handle the unstructured scenario. The contributions in this dissertation successfully address all the five open challenges mentioned at the beginning of this paragraph. We extensively test on several real world video data sets involving both single and multiple object anomalies. Results show marked improvements in detection of anomalies in both supervised and unsupervised cases when using the proposed sparsity models.