High-Performance Aerial Object Detection: Leveraging YOLO and PySpark in a Distributed Computing Environment

Résumé fourni par la source

Aerial object detection from Unmanned Aerial Vehicles (UAVs), satellite arrays, and airborne remote sensing platforms is critical for urban traffic surveillance, disaster response, precision agriculture, and environmental monitoring. However, processing gigapixel aerial imagery and high-frame-rate 4K/8K video streams introduces major computational hurdles: tiny target object scales (often occupying fewer than 16x16 pixels), arbitrary target orientations, extreme aspect ratio variances, and severe GPU memory exhaustion when downsampling high-resolution frames. Single-node deep learning workstations fail to satisfy real-time throughput requirements for large-scale drone swarms. This paper presents a high-performance, distributed aerial object detection architecture combining You Only Look Once (YOLOv8/YOLOv9) with Apache PySpark and Slicing Aided Hyper Inference (SAHI). We design a distributed dataflow pipeline that partitions ultra-high-resolution aerial frames into standardized overlapping tiles, distributes tiled inference tasks across a cluster of GPU worker nodes via PySpark Resilient Distributed Datasets (RDDs) and vectorized user-defined functions (Pandas UDFs via Apache Arrow), and reconstructs global object boundaries using Non-Maximum Suppression (NMS). Evaluated across standardized aerial benchmarks (VisDrone2023 and DOTA-v2.0), our distributed architecture achieves an mAP@0.5 score of 88.6% on tiny vehicle targets (a 17.4% improvement over standard downsampled YOLO), while scaling near-linearly across a 16-node PySpark GPU cluster to achieve an aggregated ingestion throughput of 595 frames per second (FPS).

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.