DMU-YOLO:机载视觉的多类异常行为检测算法  

DMU-YOLO:Multi-Class Abnormal Behavior Detection Algorithm Based on Air-Borne Vision

作　　者：韩佰轩 彭月平[1] 郝鹤翔 叶泽聪 HAN Baixuan;PENG Yueping;HAO Hexiang;YE Zecong(School of Information Engineering,Engineering University of PAP,Xi’an 710086,China)

机构地区：[1]中国人民武装警察部队工程大学信息工程学院,西安710086

出　　处：《计算机工程与应用》2025年第7期128-140,共13页Computer Engineering and Applications

基　　金：国家部委基金;工程大学应用研究推进项目(WYY202304);大学科研创新团队课题(KYTD202306)。

摘　　要：针对无人机航拍图像的检测算法中存在小目标识别精度低和特征提取能力不足的问题,设计了一种改进YOLOv9的多类别异常行为检测算法。该算法在模型头部加入改进的维度感知选择性集成模块,进行了有效的通道分割和融合策略,并在主干部分添加多维协同注意力机制,同时引入最大特征池化,强化了针对自建数据集的特征提取能力,而后将通用倒置残差模块与原网络的特征提取模块融合,形成了UIB-RepELAN特征提取模块,有效提升了模型检测的鲁棒性,针对难易样本不均匀分布导致的数据集长尾分布等问题,采用数据增强方法对异常类别样本进行扩充,并使用Focaler-IoU对损失函数进行重构,提高模型泛化能力。结果表明,相较于基线模型,在Vis-Drone2019数据集上的检测精度由0.046提高到0.048;针对自建数据集的检测精度由0.909提高到0.960,平均检测用时为28 ms,满足了高效率高精度的检测要求。Aiming at the problems of low accuracy of small target recognition and insufficient feature extraction ability in the detection algorithm of UAV aerial images,a multi-category anomaly detection algorithm based on the improved YOLOv9 is designed.The algorithm first adds an improved dimension-sensitive selective integration module to the model head,conducts effective channel segmentation and fusion strategies,and adds a multi-dimensional collaborative attention mechanism to the main body part,while introducing the maximum feature pooling to enhance the feature extraction ability for the self-built dataset.Then,the universal inverted residual module is fused with the original network’s feature extraction module to form the UIB-RepELAN feature extraction module,effectively improving the robustness of the model’s detection.Finally,data augmentation methods are used to expand the anomaly category samples with uneven distribution of easy and difficult samples,and the loss function is reconstructed using Focaler-IoU to improve the model’s generalization ability.The results show that the detection accuracy on the VisDrone2019 dataset is improved from 0.046 to 0.048 compared with the baseline model,and the detection accuracy on the self-built dataset is improved from 0.909 to 0.960,with an average detection time of 28 ms,meeting the requirements of high efficiency and high accuracy detection.

关 键 词：YOLOv9算法 多类异常行为检测 特征提取 无人机航拍数据集 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]