用于快速反无人机系统测试与部署的可信计算机视觉

Stanimir Arnaudov, Peter A. Chua, Thomas Mühlenstädt, Hanwool ParkResaro peter.chua@resaro.ai,thomas.muehlenstaedt@resaro.ai,hanwool.park@resaro.ai ABSTRACT Computer vision (CV) is increasingly deployed across NATO militaries for critical defense tasks likedrone detection and classification, drone tracking, and terminal guidance. This continues to revolutionizemilitary operations on the battlefield, offering a new way to keep up with overwhelming adversarycapabilities in low-cost drones employed en masse, amidst the complexities of modern urban warfare,declining manpower, and an overwhelming influx of imagery sources. Recent advancements in CVincluding more performant models with larger training datasets, and increased edge computing capacityon sensor platforms, have lowered the barrier to entry for automation. However, ethical considerations, Ensuring the reliability of CV models through robust Testing, Evaluation, Verification, and Validation(TEVV) is paramount for effective deployment. However, the limited generalization of classical CNNsacross diversereal-world conditions hinders scalability. Current TEVV methods, including publicbenchmarks, often lack relevance for specific counter-drone applications and imagery, while bespoketestingis resource-intensive.Bridging the gap between governance,operational,and technical To address these issues, we propose an assurance framework tailored for mission-oriented, high-riskdrone defense CV systems for electro-optical and infra-red imagery. This framework lays out the criteriato create robust internal reference datasets and quantify their properties, including an approach to qualifyhard-to-detect targets, and introduces comprehensive pre-deployment testing checklists for CV models.Some examples of actual implementation will be shown, as well as how highly contextualized, use-casespecific internal benchmarks can be developed to compare models in a standardized manner. Overall, thisframework aims to simplify and scale automated testing, identifying weak performance spots, and 1.0INTRODUCTION Computer Vision (CV) is a field of study relating to how computers can understand and analyze informationfrom the environment using visual data such as images and videos, across a wide range of imagingtechniques and platforms including electro-optical, thermal, and radar sensors. Over the past decade, CV has Assuring Trustworthy Computer Visionfor Rapid Counter-Drone System Testing and Deployment made significant advancements from the increasing adoption of techniques in artificial intelligence (AI)including machine learning, deep learning, and pattern recognition. Accordingly, CV systems have also seenwidespread use in defense applications and systems, augmenting human operators in processing andanalyzing data from a wide variety of imaging sensors for target detection, classification, tracking, andprosecution. Applied to the counter-drone defense mission, CV systems can be useful in (a) extendingoperators’ ability to perceive small-sized drones beyond relying on human sight, (b) ease classification,tracking and prosecution tasks through computer automation and therefore handle more challenging targets,(c) compensate for the other counter-drone sensors’ limitations and enhance system redundancy, at the 2.0RATIONALE The rapid proliferation of drones across the inventories of potential adversaries, spurred by advances incommercial and recreational applications but adapted for military uses, has concurrently given rise to acritical challenge: the need for effective counter-drone measures. This escalating drone adoption has spurredthe development of numerous solutions aimed at detecting, tracking, and neutralizing unauthorized or hostiledrones. However, the sheer variety and complexity of these emerging technologies present significanthurdles for decision-makers tasked with their deployment. The process of selecting the most appropriate Crucially, decision-makers must also anticipate and understand the circumstances under which their chosensystems might perform sub-optimally or even fail entirely. Simon Burton and Benjamin Herd from theFraunhofer Institute for Cognitive Systems in Germany note that “the role of safety assurance can be seen asstriving to facilitate decisions of type II wherever type I is not possible, whilst avoiding type III decisions”[2], in reference to Knight’s “three types of decisions: decisions under certainty (type I) where theconsequences of all options are known; decisions under risk (type II) where possible futures are known,probability distributions are known, and statistical analysis is possible; and decisions under uncertainty (typeIII) where the future states are known but the probabilities are unknown”[8]. For decision-makers in defenseand security, these activities of safety assurance are also relevant to objectives outside of system safety, such Any proposed counter-drone solution, whether a standalone system or a combination of techno