Advanced classification techniques for real-time signals in resource-constrained systems

Automated classification of situational awareness data collected by autonomous vehicles is currently an unmet need in many applications. Classification algorithms developed at JHU/APL extend large margin classification (LMC) machine learning techniques to solve domain-specific problems such as those found in unmanned undersea vehicles systems. Common classification issues for the described systems include the following: (1) Asymmetric binary class membership, that is, a small amount of a target signal must be distinguished in a very large collection of data; (2) no silver-bullet features, i.e., robust classification requires many weak features, are used to distinguish targets from other signals; and (3) limited processing resources. The JHU/APL solution uses existing LMC technology with modifications to solve specific domain issues: (1) addition of a penalty term to address class asymmetry, (2) an iterative training algorithm that yields a sparse solution optimized for a given computational footprint, and (3) featureless classification that requires minimal or no data reduction and can improve classification robustness of the algorithm and cut development costs. Use of these developed techniques is demonstrated using well studied open source data that are representative of that required for autonomous vehicle classification tasks.