Memory-Efficient Annotation Techniques for Autonomous Drone Navigation

The rapid adoption of autonomous drones in mission-critical applications such as infrastructure inspection and emergency response demands highly accurate real-time (RT) environmental perception, yet faces fundamental bottlenecks in its perception system due to memory-intensive data representations. We present a novel end-to-end framework that synergistically unifies: (1) dynamic scene complexity-aware perception-mode selection, (2) ASPP-optimized sparse mask encoding achieving >90% memory reduction, (3) seamless Datumaro integration for version-controlled, AES-256 encryption data management with just 9.1ms overhead. Our hybrid bounding-box/semantic segmentation (BB/SS) switching mechanism dynamically adapts annotation granularity based on RT scene complexity, achieving notable memory-accuracy trade-off for edge-deployment.The holistic system demonstrates considerable performance through large-scale benchmarking across seven standardized data formats (JSON/CSV/Parquet/HDF5/YAML/Avro/SQLite3) and 90+ reference methods:With an energy efficiency of 10.4 FPS/W on NVIDIA hardware, achieving 92.5% mIoU (outperforming comparable lightweight methods by 23.6%) at a 1.1GB memory footprint - a 2.3Ö throughput gain over fragmented approaches enabled by TensorRT-optimized execution (56% latency reduction). These advances optimize three competing factors: RT performance (sub-5.22 ms storage latency), memory efficiency (1.1GB footprint), and perception accuracy (92.5% mIoU), demonstrating significant improvements forUAVperception pipelines as quantified in our systematic comparison against 24 reference systems. Future research will explore explainable AI (XAI), adaptive annotation, and hybrid storage to extend this integrated framework’s capabilities for nextgen aerial mobility applications.