UAV-YOLO Model v1

This research is based on the hypothesis that object detection performance in UAV-based traffic management systems can be significantly improved by redesigning the YOLOv5s architecture to better address the unique challenges of aerial imagery. These challenges include extremely small object sizes, large scale variation within a single frame, frequent occlusion, and limited onboard computational resources. The study proposes that integrating lightweight convolutional modules, Transformer-based attention mechanisms, enhanced multi-scale feature fusion, and an additional detection head into YOLOv5s can improve detection accuracy particularly for small and densely distributed objects while maintaining real-time inference capability suitable for UAV deployment. To evaluate this hypothesis, a dataset consisting of UAV-captured aerial images and video frames was employed. The data represents realistic traffic environments such as highways, intersections, and urban road networks, containing vehicles, pedestrians, bicycles, and other road users observed from varying altitudes and camera angles. The dataset was sourced from publicly available UAV traffic datasets and further supplemented with additional annotated aerial imagery to ensure diversity in object categories, environmental conditions, object densities, and scale variations. All images were manually annotated using bounding boxes following standard object detection protocols, with particular attention given to small and partially occluded objects that are critical for evaluating UAV-based detection performance.

本研究基于如下假设：通过重新设计YOLOv5s架构，可显著提升基于无人机（UAV）的交通管理系统中的目标检测性能，以更好地应对航空影像特有的挑战。此类挑战包括：目标尺寸极小、单帧内尺度差异极大、频繁出现遮挡，以及机载计算资源有限。本研究提出，将轻量级卷积模块（lightweight convolutional modules）、基于Transformer的注意力机制（Transformer-based attention mechanisms）、增强型多尺度特征融合模块（enhanced multi-scale feature fusion）以及额外检测头（additional detection head）集成至YOLOv5s中，可在保持适配无人机部署的实时推理能力的同时，提升检测精度，尤其针对小型且密集分布的目标。为验证该假设，本研究采用了由无人机拍摄的航空影像与视频帧组成的数据集。该数据集涵盖高速公路、交叉口、城市道路网等真实交通场景，包含从不同高度与拍摄角度采集的车辆、行人、自行车及其他道路参与者。该数据集源自公开可用的无人机交通数据集，并补充了额外的标注航空影像（annotated aerial imagery），以确保目标类别、环境条件、目标密度及尺度变化的多样性。所有影像均遵循标准目标检测协议（standard object detection protocols），采用边界框（bounding boxes）进行人工标注，并重点关注对无人机目标检测性能评估至关重要的小型及部分遮挡目标。