Certified ML Object Detection for Surveillance Missions

TL;DR

The paper addresses certifying ML-based object detection for drone surveillance by outlining a partial ARP 6983–compliant development workflow. It defines an ODD and an ML CODD (MLCODD), builds bias-aware datasets via augmentation and mosaic tiling, and designs a tiling-based YOLOv3-tiny detector optimized for real-time execution on NVIDIA Xavier AGX. It demonstrates end-to-end implementation, including CPU and GPU code paths with traceability, and introduces a custom GEMM approach and FP16 quantization to meet latency and memory constraints. The work provides a practical blueprint for integrating ML components in safety-critical surveillance systems and identifies concrete future steps toward fuller ODD definition, complete certification evidence, and system-level monitoring.

Abstract

In this paper, we present a development process of a drone detection system involving a machine learning object detection component. The purpose is to reach acceptable performance objectives and provide sufficient evidences, required by the recommendations (soon to be published) of the ED 324 / ARP 6983 standard, to gain confidence in the dependability of the designed system.

Figures (14)

• Figure 1: Surveillance Area
• Figure 2: ARP 6983 simplified development workflow (adapted from gabreau:hal-03761946)
• Figure 3: ML Constituent
• Figure 4: Object positions in the original dataset on the left and the augmented dataset with unbiased position on the right
• Figure 5: Distribution of size of object in the dataset