THIS IS A CONTINUATION OF N00014-14-1-0484 Deep Structures Boosted Self-Organized Behavior Pattern Learning for Anomaly Detection
Abstract
Statement of Work:Develop learning algorithms for detecting anamolous behavior in video.Objective:Develop learning algorithms for detecting anamolous behavior in video.Approach:Representing and detecting abnormal activities--the atypical behaviors of single individuals or a crowd of people--in surveillance imagery are fundamental and challenging research areas with important applications in defense and security. Examples of such activities include loitering, intrusion or trespassing, walking in the wrong direction, vandalism, unauthorized gatherings, street fights, crowd stampede, etc. Open challenges in this field include unclear definition of anomaly, non-trivial learning semantics and contexts from diverse low-level features, and tedious manual annotations from large-scale and high-dimensional videos. Dr. Fu s proposed research is to systematically and rigorously address the basic research challenges of anomaly detection by creating a new machine learning frameworkthat interweaves deep-learning architectures with low-rank constraints and self-organized behavior pattern learning for robust and real-time video surveillance. Dr. Fu will (a) investigate performance of deep-learning structures; (b) formulate visual representations under real-world conditions in the presence of realistic uncertainties; and (c) leverage and extend the design of future visual computational systems by novel machine learning methods that can process large-scale, multi-source, multi-label data, allowing for applications to operate at their optimal performance levels.Overall Merit and ONR Mission/Relevance:This effort is expected to develop novel and effective machine learning methods for understanding activities in video, in particular methods for representation and detection of anamolous behaviors. This effort is directly related to the Naval Focus Areas of Autonomy and Information Dominance. This effort will advance representions and detecting of abnormal activities the atypical behaviors of single individuals or a crowd ofpeople in surveillance imagery which are important applications in defense and security.ProgressThe PI~s proof-of-concept study has shown that the originally proposed deep structures boosted pattern learning is very effective for dealing with anomaly detection challenges. They have been developing deep low-rank coding models and deep linear coding models, and further integrated such techniques with sparse graph, tensor representation, manifold learning, graph embedding, subspace learning, multi-view learning, and transfer learning for generalized frameworks. These new techniques have been applied to outlier detection, which is the abstract scenario of anomaly detection. They evaluated these methods in a number of applications including video segmentation, human re-identification, RGBD action recognition, face recognition, action prediction, and visual clustering, etc. Technical outcomes have been sufficiently demonstrated to be effective and efficient which can be tailored to real-world environments. The results were published in 28 papers, and the PI and his group received 9 awards/honors.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 23, 2016
- Source ID
- N000141612794
Entities
People
- Raymond Fu
Organizations
- Northeastern University
- Office of Naval Research
- United States Navy