By 2020, it is projected that every human being will contribute 1,000GB of the sensor data individually. These sensors include but not limited to GPS, accelerometer, gyroscope, microphone, camera, all kinds of wearable biomedical sensors, etc. Beyond the domain of monitoring human activity, multi-sensor data processing has been an active research topic within the context of numerous practical applications, such as medical image analysis, remote sensing, and military target/threat detection. One powerful tool to tackle these critical Big Data problems is sparsity-driven signal-processing techniques. A sparse representation not only provides better signal compression for bandwidth/storage efficiency, but also leads to faster processing algorithms as well as more effective signal separation for detection, classification and recognition purposes since it focuses on the most intrinsic property of the data. Sparse signal representation allows one to capture the hidden simplified structure present in the data jungle, and thus minimizes the harmful effects of noise in practical settings.
This research investigates the problem of dynamic dictionary learning to obtain the most effective and adaptive sparse data representation even in the presence of low-rank interference. The investigator seeks to develop a general framework that takes advantage of diversity, yet complementary, features in vast correlated data sources collected from multiple, possibly heterogeneous, sensors co-located in the same spatio-temporal physical space, recording the same physical event. Such scenarios ensure that interference noise patterns are very similar, hence justifying the low-rank property of the interference, while allow specific structural sparsity pattern of the signal of interest to be captured in the dictionary and/or in the sparse codes.
