DESIGN OF CNN DETECTORS FOR SMALL-SIZED UAVS FOR PERIPHERAL DEVICES TO ACHIEVE A COMMITMENT BETWEEN ACCURACY AND SPEED

Abstract

This article investigates the application of convolutional neural networks (CNNs) to improve the accuracy and robustness of small unmanned aerial vehicle (UAV) detection in images and video streams, with a particular focus on one-stage real-time detectors derived from the YOLO family and their lightweight adaptations for edge deployment. The study outlines how preserving fine-grained spatial cues during downsampling, strengthening multi-scale feature fusion in the neck/head, and incorporating selective (cost-aware) attention modules can enhance the detection of tiny targets while reducing false alarms caused by birds, clouds, compression artifacts, and cluttered backgrounds. It examines the core design mechanisms of CNN-based detectors for small objects, emphasizing the role of high-resolution branches, efficient feature pyramid topologies, and stable bounding-box regression when objects occupy only a few pixels. Additionally, the article discusses key evaluation aspects, including the importance of small-object metrics and the speed–accuracy trade-off that governs practical anti-UAV systems operating under strict latency, memory, and power constraints. Furthermore, the article considers operational challenges such as domain shifts across landscapes and weather conditions, low-light and infrared scenarios, and the need for temporal consistency in video, where integrating post-processing and tracking can improve stability beyond frame-level performance. By synthesizing recent research trends and practical constraints, the article underscores the necessity for continued development of CNN-centric design strategies and benchmarking protocols to support reliable real-time UAV detection on resource-limited platforms. To better handle extremely small targets, it highlights low-overhead choices such as a P2 scale branch, anti-aliasing downsampling, IoU-aware losses that stabilize regression on tiny boxes, calibrated confidence scoring, hard-negative mining, and deployment-minded quantization/pruning with NPU-friendly operators, etc.