Challenge

Photo Courtesy of U.S. Army

The success of military objectives and the safety of military personnel depend on the ability to detect, classify, and identify harmful objects of interest, such as military vehicles, land and sea mines, and improvised explosive devices (IEDs). Since World War II, mines have damaged more U.S. Naval vessels than all other threats combined. Roadside bombs and other IEDs are the single greatest threat to currently deployed coalition forces, and IEDs have been responsible for almost 40% of U.S. casualties in Operation Iraqi Freedom.

For decades, the U.S. military has been integrating advanced radar and sonar sensor systems on airborne and undersea platforms to aide in the detection, classification, and identification of air, ground, and underwater targets. Recent growth in the development and deployment of unmanned systems, including unmanned aerial vehicles (UAVs) and unmanned undersea vehicles (UUVs), has greatly expanded the utility of multiple sensor modalities (radar, sonar, EO/IR) and the quantity of collected sensor data. Human analysts are often tasked with manually examining sensor data. However, the overwhelming volume of data exceeds the limited available human resources, and performance is often unreliable due to human fatigue and distraction. Therefore, automated real-time sensor processing techniques are required to reliably detect and discriminate targets of interest while minimizing the number of false alarms. The success of automated techniques depends on the sensor system, the observed background objects, and environmental conditions, all of which can influence the performance of automated systems.

Benefits of SIG’s Approach

SIG has developed advanced statistical data modeling and inference techniques that accurately and efficiently detect and discriminate targets in sensor data. By maintaining a probabilistic framework, SIG’s predictive modeling capabilities produce more meaningful and useful results than typical discrimination strategies. Model uncertainty can be quantified to determine if additional data is necessary, decision-making is aided by confidence values associated with target declarations, and model components and values can be probabilistically combined. As data is collected, SIG’s techniques are capable of adapting to changes in sensors and environmental conditions, as well as exploit all available data regardless of whether object identifications are known.

Products/Solution

SIG’s target recognition software suite offers a combination of state-of-the-art detection, classification, and identification techniques that can be tailored to any application space. The software contains multiple statistical data models within a flexible and modular architecture that is readily optimized for any sensor modality, including the fusion of multiple sensors. The software can be configured for onboard integration with manned or unmanned platforms to provide real-time processing or delivered as stand-alone modules to support analysts and post-mission objectives.

Technology Summary

SIG’s target detection, classification, and identification algorithms are rooted in Bayesian statistical principles applied to multiple technical disciplines, including machine learning, pattern recognition, and statistical signal processing. The predictive modeling technology utilizes prior probability distributions and likelihood functions designed to optimally map sensor data to a target classification. Both linear and nonlinear models are utilized in supervised, semi-supervised, and unsupervised machine learning algorithms for class discrimination and density estimation.