Hybrid Architecture for Real-Time Object Detection from UAV Video Streams Using External ROI Processing and Embedded Optimization

This paper presents a hybrid system architecture designed to enable real-time object detection from unmanned aerial vehicle (UAV) video streams by leveraging both external and embedded processing resources. Due to the computational limitations of onboard platforms such as the Raspberry Pi 5, the system offloads deep learning-based detection tasks to an external GPU-equipped server that applies YOLOv8 to extract Regions of Interest (ROI). These ROIs are transmitted via MQTT and subsequently classified using a lightweight MobileNet SSD model on the embedded device. Experimental evaluations involving RGB and thermal video data reveal that the ROI-based approach reduces processing load and improves temporal detection consistency in well-lit and dynamic scenarios. However, challenges such as increased false positives and performance degradation in thermal imagery highlight the need for adaptive strategies and further optimization. The proposed architecture demonstrates a viable pathway for achieving efficient, real-time surveillance in maritime and resource-constrained environments.