摘要
为解决夜间场景下视频监控目标检测在实际应用时准确率不高这一问题,提出改进的YOLOv5算法。首先,建立了真实夜间场景目标的数据集,该数据集有2000张图像,分为了机动车、非机动车和车牌三个类别,以8∶2的比例均匀随机分为训练集和测试集,将夜间目标的图像放入改进的YOLOv5模型中训练,最终达到在夜间检测目标的目的;改进的YOLOv5利用了K-means++聚类算法生成自适应锚框,提高对夜间目标样本的聚类效率。其次,将改进的CBAM注意力机制与特征提取网络进行融合以获取夜间目标的重要特征。最后,将Bottleneck替换成GSBottleneck模块,利用GSConv轻量化的优势减少网络模型的计算量与参数量。结果表明,通过原YOLOv5网络模型算法训练后得到的mAP值为86.69%,改进后的YOLOv5网络模型算法训练后得到的mAP值为91.98%,三种被检测类别:机动车、非机动车和车牌的检测准确精度与原版算法相比分别提升了2.00、6.66、7.19个百分点,改进的YOLOv5网络模型可以为夜间场景下车辆特征的检测提供较好的技术支持。
To solve the problem of low accuracy in practical application of video surveillance target detection at night,here an improved yolov5 is proposed.A dataset of real night scene is established,this dataset was put into the improved YOLOv5 for train-ing,the purpose of detecting targets at night was achieved.To improve the clustering efficiency of target samples at night,in the im-proved YOLOv5,the K-means++clustering algorithm is utilized to generate adaptive anchor boxes.The improved CBAM attention mechanism was put into the CSP_X module to get important features of the target at night.Replace Bottleneck with the GSBottle-neck module,which is to reduce the computation and parameters of the network model.By experimental results,the mAP values af-ter training by YOLOv5 and improved YOLOv5 were 86.69%and 91.98%,respectively,after training with improved YOLOv5,the AP values of motor vehicles,non-motor vehicles and license plates increased by 2.00,6.66 and 7.19 percentage,respectively.The improved YOLOv5 can provide better technical support for detecting vehicle characteristics at night.
作者
张奕博
张雅丽
Zhang Yibo;Zhang Yai(College of Information and Cyber Securiy,People’s Public Security University of China,Beijing 100038,China)
出处
《现代计算机》
2024年第2期18-25,共8页
Modern Computer
基金
中国人民公安大学安全防范工程双一流创新研究专项(2023SYL08)。
