STAN: Scene and Target Analysis Network

The introduction of computer vision in precision-guided ammunition will potentially increase their targeting capabilities, in particular for moving targets such as tanks, vehicles and mobile command posts. EO/IR cameras offer a relatively inexpensive, low-SWaP payload solution which, combined with latest advances in SIMD hardware (GPUs), are finally starting to allow running sophisticated visual guidance algorithms onboard. Although modern computer vision may confer great benefits in targeting precision and accuracy, it also suffers from long-standing problems such as partial occlusion. We propose to tackle occlusion by leveraging advances in computational neuroscience into a recent deep neural network for target detection. In our algorithm, a part-based target model works jointly with an occlusion model. The part-based model decomposes target signatures in parts, and the occlusion model decomposes the scene in terms of occluded/occluders. By requiring the detection of an occlusion when only object part(s) are visible, the combined model will be more robust to false positives, avoiding the issue that individual parts alone are hard to distinguish from background clutter. The focus of the Phase I efforts is (1) the development of a vision-based algorithm for object detection with occlusion, and (2) concept demonstrations and performance characterization on representative imagery.