Lightweight classification models such as the LDDm-CNN have shown strong potential for distinguishing drones from other aerial objects such as airplanes. However, classification alone is insufficient for real-time surveillance, since it requires prior cropping or manual region-of-interest extraction and cannot localize drones directly within a scene. Although YOLO-based detectors achieve remarkable mAP, their large model sizes and high computational demands hinder deployment on edge devices. Therefore, this renders the architecture unsuitable for real-time drone surveillance in resource-constrained environments. This limitation reduces applicability in dynamic security environments where both recognition and spatial localization are critical. To overcome this gap, we extend the LDDm-CNN into a full detection framework optimized for edge devices. The proposed LDDm-YOLO uses the LDDm backbone as a compact feature extractor and integrates a lightweight, anchor-free detection head with a shallow feature pyramid for multi-scale object localization. Knowledge distillation transfers rich spatial and semantic features from a larger teacher detector, while Bayesian optimization tunes key hyperparameters such as distillation temperature, loss weighting, and feature fusion depth. Experiments on public drone detection datasets show that LDDm-Det achieves competitive mean Average Precision (mAP =0.95) while retaining a smaller model size of only a few megabytes, enabling real-time detection on edge devices and localization on resource-constrained devices.