基于跨尺度动态特征金字塔的无人机图像车辆检测算法

Vehicle Detection Algorithm in UAV Images Based on Cross-Scale Dynamic Feature Pyramid

摘要: 近年来，无人机(UAV)在交通监控和智能停车等多个领域得到了广泛应用，其中车辆的实时监测和分类成为其关键任务之一。车辆检测面临多种挑战，尤其是在小型车辆和无人机飞行角度变化引起的目标尺度变化下，检测网络优化的难度加大。此外，高空航拍图像中的小目标使得可提取的特征有限，进一步影响检测精度。为了解决这些问题，本文基于YOLOv8算法提出了一种高效且实时的车辆检测网络，主要改进包括：1) 在网络的backbone部分引入CPCA注意力模块，以增强模型对小目标的关注能力，进而提升特征提取效果；2) 对YOLOv8的Neck结构进行改进，借鉴DAMO-YOLO中的GFPN思想，以较小的参数量显著提升了检测精度，同时将传统的双线性插值上采样替换为DySample动态上采样，使模型能更好地适应目标尺度变化，最终构建了Cross-Scale Dynamic Feature Pyramid Network (CS-DyFPN)网络；3) 提出了Inner-Focaler-IoU损失，结合了Inner-IoU与Focaler-IoU的优势，能够自适应地聚焦困难样本，相比CIOU提升了检测精度。实验结果表明，本文方法在VisDrone2019数据集上相较于原始YOLOv8算法，在实时性和准确性方面取得了显著提升，特别是在小目标检测任务中表现优异。

Abstract: In recent years, unmanned aerial vehicles (UAVs) have been widely applied in various fields, such as traffic monitoring and smart parking, where real-time vehicle detection and classification have become critical tasks. Vehicle detection faces several challenges, particularly due to target scale variations caused by small vehicles and changes in the flight angle of drones, which complicate network optimization. Additionally, small targets in aerial images limit the features that can be extracted, further affecting detection accuracy. To address these issues, this paper proposes an efficient and real-time vehicle detection network based on the YOLOv8 algorithm. The main improvements include: 1) Introducing the CPCA attention module into the backbone of the network to enhance the model’s focus on small targets, thereby improving feature extraction; 2) Modifying the Neck structure of YOLOv8, inspired by the GFPN concept from DAMO-YOLO, which significantly improves detection accuracy with fewer parameters. Additionally, the traditional bilinear interpolation upsampling is replaced by DySample dynamic upsampling to better adapt to target scale variations, resulting in the Cross-Scale Dynamic Feature Pyramid Network (CS-DyFPN); 3) Proposing the Inner-Focaler-IoU loss, which combines the advantages of Inner-IoU and Focaler-IoU, allowing the model to focus on difficult samples and improving detection accuracy compared to CIOU. Experimental results show that the proposed method significantly improves both real-time performance and accuracy on the VisDrone2019 dataset, particularly excelling in small target detection tasks compared to the original YOLOv8 algorithm.