Kitware End-to-end Recognition, Modeling, and Inference of Trajectories (KERMIT)

We are proud to propose Kitware End-to-end Recognition, Modeling, and Inference of Trajectories (KERMIT), an SBIR project to develop highly accurate target detection, tracking, classification, and motion prediction algorithms for fast-moving targets captured in overhead persistent infrared (OPIR) imagery from satellites in low earth orbit (LEO). The proposed research builds on Kitware’s extensive history of OPIR detection and tracking, and will leverage an existing OPIR simulator to both enable rapid experimentation and reduce the risk of performance losses when transitioning to real OPIR data in later phases. Our approach’s state-of-the-art elements include the co-optimization of sensor configuration with subsequent processing, the incorporation of a transformer network to refine initial high-recall detections, and the use of pseudo-3D network architectures to improve accuracy while reducing memory and processing requirements, in support of future low-SWaP (Size, Weight, and Power) deployment.