Confabulation Based Real-time Anomaly Detection for Wide-area Surveillance Using Heterogeneous High Performance Computing Architecture.

摘要

The feasibility of probabilistic inference based anomaly detection was determined, and those results were applied to wide area surveillance. An abstract - level autonomous information processing framework was developed that provided continuous monitoring and real - time anomaly detection over hundreds of square kilometer areas. The anomaly recognition and detection (AnRAD) system was built as a cogent confabulation network. It represented road traffic using a set of features extracted from a Ground Moving Target Indicator (GMTI) input stream and performed likelihood-ratio testing on a set of key features to detect abnormal vehicle behavior. Due to its low learning and recall complexities, the AnRAD supported incremental learning, which was proved to enhance the detection accuracy. A self - structuring technique was developed that learned the structure of a probabilistic inference network from unlabeled data. Without any assumption of the distribution of data, mutual information between features was leveraged to learn a succinct network configuration. Compared to several existing anomaly detection methods, the proposed approach provided higher detection performances and excellent reasoning capabilities. Massive parallelism was inherent to the inference model and accelerated the detection process using state- of-the-art multicore processors including graphic processor units (GPUs) and Intel Xeon Phi processors. Experimental results showed significant speedups, which can enable real-time applications with high-volume data streams.