一种基于改进YOLOv7的落水人员检测方法  

A detection method for drowning personnel based on improved YOLOv7

作　　者：黄旭 施闰虎[2] 曾孟佳 Huang Xu;Shi Runhu;Zeng Mengjia(School of Electronic Information Huzhou College,Huzhou 313000,China;School of Information Engineering Huzhou University,Huzhou 313000,China;Huzhou Key Laboratory of Urban Multidimensional Perception and Intelligent Computing,Huzhou 313000,China)

机构地区：[1]湖州学院电子信息学院,湖州313000 [2]湖州师范学院信息工程学院,湖州313000 [3]湖州市城市多维感知与智能计算重点实验室,湖州313000

出　　处：《现代计算机》2024年第15期25-31,共7页Modern Computer

基　　金：教育部人文社会科学一般项目(20YJCZH005);浙江省湖州市工业攻关项目(2018GG29)。

摘　　要：溺水事故一旦发生,对溺水对象的及时发现并采取积极的相应措施,在减小伤亡后果方面有着至关重要的作用,因此对落水人员的自动监测变得尤为重要。在对野外落水人员目标进行模拟检测的过程中发现,自然水域场景复杂,具有检测目标小、环境干扰大等问题,现有的基准模型YOLOv7的目标漏检率较高。通过在YOLOv7模型的骨干组件中加入ResNet-ACmix,以及在Head部分增加ACmix,从而保留了Backbone采集到的特征,有效提取了复杂场景下的落水人员小目标的特征信息,并增强了模型对落水人员小目标的特征感知和位置信息,降低了水域复杂环境对特征提取的干扰,改进SPPCSPC中的池化部分,保证了落水人员小目标在大场景复杂水域的定位,进一步降低环境干扰,并提升检测速度。在构建的落水人员数据集上进行的各种实验表明,与基线网络的YOLOv7算法相比,改进的YOLOv7平均精度达到80.7%,比原始网络提高了7%。消融实验表明,所设计的模块可以提高检测精度,并能直观地显示不同场景下的检测效果。实验验证了改进YOLOv7在落水人员目标检测中的适用性。Once a drowning accident occurs,timely detection of the drowning object and taking active corresponding measures play a crucial role in reducing the consequences of casualties.Therefore,automatic monitoring of drowning personnel has become particularly important.In the process of simulating the detection of personnel falling into the water in the wild,it was found that the natural water scene is complex,with problems such as small detection targets and large environmental interference.The existing benchmark model YOLOv7 has a high target miss rate.A ResNet ACmix has added to the backbone component of the YOLOv7 model,and an ACmix has added in the Head section to preserve the features collected by Backbone.It effectively extracts the feature information of small targets of underwater personnel in complex scenes,enhances the model’s feature perception and position information of small targets of underwater personnel,reduces the interference of complex water environments on feature extraction,and improves the pooling part in SPPCSPC.It ensures the positioning of small targets for drowning personnel in large and complex water areas,further reducing environmental interference and improving detection speed.Various experiments conducted on the constructed dataset of drowning personnel have shown that compared with the YOLOv7 algorithm of the baseline network,the improved YOLOv7 algorithm has an average accuracy of 80.7%,an increase of 7%.The ablation experiment shows that the designed module can improve detection accuracy and visually display the detection results in different scenarios.The experiment verified the applicability of improved YOLOv7 in target detection of underwater personnel.

关 键 词：YOLOv7 落水人员 注意力机制 小目标检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]