Comparative Analysis of State-of-the-Art Object Detection Frameworks for Real-Time UAV Applications in Urban Environments

Abstract

Object detection, a critical task in computer vision, has significant applications in autonomous driving, surveillance, and Unmanned Aerial Vehicle (UAV-based) monitoring, where accurate identification and localization of objects are vital. This paper performs an analysis on the state-of-the-art object detection models and evaluates their effectiveness in detecting “person,” and “vehicle,” classes in real-time applications. A subset of the Common Objects in Context 2017 dataset with over 1,000 annotated images was used by the application of pre-trained models, namely Faster R-CNN, You Only Look Once (YOLO) of v5 and v8, Mask R-CNN, EfficientDet, and SSD. The novelty lies in introducing a new set of parameters for YOLOv8 that optimized its performance for UAV-based urban monitoring tasks. A testing was conducted across 39 model variants from TensorFlow Hub and assessed based on precision, recall, and F1-score. YOLOv8 achieved the highest F1-score of 0.8 for the “person” class, while Faster R-CNN ResNet152 V1 gave balanced performance, suitable for speed-accuracy trade-offs applications. Real-time testing on urban road images validated YOLO's superiority for time-sensitive tasks and highlighted needs for advanced tracking and multi-frame future analysis.