METHODOLOGY AND APPLICATIONS OF AI ENABLED IMAGE RECOGNITION IN UAVS
Balayev, F.R., Hasanov, Arif, Hashimov, Elshan G. · eNTUKhPIIR Repository (Kharkiv Polytechnic Institute) · 2025
This paper provides a systematic examination of artificial intelligence integration for image recognition on unmanned aerial vehicles, covering methodology, application domains, and operational implications.The motivation is the need to secure information superiority by reducing the sensing to decision latency while maintaining accuracy and resilience in contested environments.We propose an end-to-end Edge AI pipeline that runs entirely on the platform and combines multisensor inputs with optimized deep learning models.The architecture comprises five sequential stages: sensor acquisition from EO IR and SAR payloads, preprocessing for compression and normalization, object detection and tracking using configurable families such as YOLO, Faster R CNN, and DETR paired with DeepSORT, human in the loop validation for accountability, and mission management that route confirmed detections to navigation and payload subsystems.The design emphasizes real time operation on GPU or NPU based system on chip modules with strict energy budgets.Dataset construction follows three principles.First, domain relevance through acquisition in mountainous and urban terrains at varied altitudes and illumination.Second, interoperable labeling using COCO and VOC formats with classes such as armored vehicles, artillery, air defense, logistics vehicles, and personnel.Third, robust splits and class balance using a 70 15 15 partition with focal loss and targeted relabeling when necessary.Performance is evaluated by four axes.Accuracy uses mAP@[0.5:0.95],F1 score, false alarm rate, and ID switch count for tracking.Speed is measured as end-to-end latency and frames per second with a target below 50 ms total latency.Energy efficiency is quantified as joules per frame, where quantization and pruning yield 20 to 35 percent savings.Reliability is assessed under fog, night, urban clutter, camouflage, communications degradation, and GNSS spoofing.Experimental results demonstrate the inherent speed accuracy trade off.A quantized YOLO S achieves roughly 35 FPS with mAP near 0.55, YOLO M improves accuracy to about 0.62 at 22 FPS, while Faster R CNN reaches mAP near 0.64 with only 8 FPS.These findings support a configurable model portfolio where mission requirements and onboard compute define the operating point.Multisensor fusion at the AI level reduces false positives by 18 to 27 percent when EO IR and SAR are jointly exploited, which is especially beneficial at night and in low visibility.A representative latency profile allocates 8 to 12 ms to sensor and preprocessing, 15 to 28 ms to inference, and 5 to 10 ms to post processing, remaining under the 50 ms threshold.Thermal headroom in hot climates emerges as a practical constraint that motivates improved heat sinking and dynamic frequency scaling.