改进YOLOv7算法的工业安全多目标检测研究

Research on the Improvement of YOLOv7 Algorithm for Industrial Safety Multi-Target Detection

在工厂阀室内,安全问题一直是一项关注的重点,为了防止工厂阀室内员工未佩戴安全帽、阀室内抽烟、打电话分心等异常行为对工厂设备和员工造成伤害,目标检测技术常常被应用于工业安全场景下进行员工异常行为的识别,从而保证工业生产安全进行。目标检测技术作为工业安全监控的核心组成部分,其在真实场景下应用、算法识别性能尤为关键。为了进一步研究应用于工业生产安全领域的人体异常行为识别算法,本文进行了如下工作:(1) 本文提出了一个近两万张工厂阀室背景下的数据集,数据集共2万多张真实工厂阀室下的监控拍摄图片。将其按照训练集(验证集)和测试集按照4:1的比例进行划分。为了数据集的实用性,对风险等级最高的8种异常行为以及三种安全行为进行数据标注形成了11类标签的USDA数据集,并且为了提高模型的鲁棒性,还对USDA数据集进行数据增强扩充,使其更具有在真实场景下应用的实用性。(2) 本文提出了改进的SD-YOLOv7模型进行异常行为的识别。在该模型中,首先在核心网络Backbone内集成了Squeeze-and-Excitation Networks (SENet)的注意力机制。其中,SENet引入了一个创新的特征重标定方法,它能够自动学习并分配各个特征通道的权重,从而提高关键特征通道的权重值,通过自学习的方式自动获取每个特征通道的权重,增大目标重要特征通道的权重值,同时加入可变形卷积DCNv4来替换传统卷积,以更好地捕捉各种角度和姿态的目标以加强SD-YOLOv7模型在工厂监控下在不同角度下的目标检测能力,此外还提出了一种符合工厂阀室场景特征的加权损失函数FA loss。本模型在自建的2万多张图像数据集下与传统YOLOv7、YOLOv9等模型做了对比实验。结果表明本模型在召回率、平均精确率(mAP)比传统YOLOv7、YOLOv9等模型有较大提升,改进后的SD-YOLOv7算法在增加较少复杂度的情况下明显提升了算法的性能。此外,该模型已经成功部署在边缘设备上,可以成功地为合作单位进行实时检测。本研究为工业安全监控领域提供了一个高度实用的数据集和一种高效的目标检测模型,未来将探索在更多实际工业场景中的应用。

Safety has always been a focal point within factory valve rooms. To prevent abnormal behaviors such as employees not wearing safety helmets, smoking within the valve room, or being distracted by phone calls from causing harm to factory equipment and personnel, target detection technology is often utilized in industrial safety scenarios for the recognition of employee abnormal behaviors, thereby ensuring the safety of industrial production. Target detection technology, as a core component of industrial safety monitoring, is particularly crucial for its application in real scenarios and the performance of algorithm recognition. To further study the human abnormal behavior recognition algorithms applied in the field of industrial production safety, this paper has conducted the following work: (1) This paper proposes a dataset of nearly 20,000 images set against the backdrop of a factory valve room, consisting of over 20,000 real surveillance photos taken in actual factory valve room settings. The dataset is divided into training (validation) and testing sets in a 4:1 ratio. To enhance the practicality of the dataset, data annotation was conducted for the 8 highest-risk abnormal behaviors and three safe behaviors, forming an 11-class labeled USDA dataset. Additionally, to improve the robustness of the model, data augmentation was applied to the USDA dataset to make it more applicable in real-world scenarios. (2) This paper introduces an improved SD-YOLOv7 model for the recognition of abnormal behaviors. In this model, the Squeeze-and-Excitation Networks (SENet) attention mechanism was first integrated into the core network Backbone. SENet introduces an innovative feature recalibration method that can automatically learn and allocate weights to each feature channel, thereby enhancing the weight of key feature channels. It automatically acquires the weights of each feature channel through self-learning, increasing the weight of important target feature channels. At the same time, deformable convolution DCNv4 is used to replace traditional convolution to better capture targets at various angles and postures, thereby strengthening the target detection capabilities of the SD-YOLOv7 model in factory monitoring under different angles. Furthermore, a weighted loss function, FA loss, was proposed that is tailored to the characteristics of the factory valve room scenario. This model was compared with traditional YOLOv7, YOLOv9, and other models on a self-built dataset of over 20,000 images. The results indicate that this model has significantly improved in recall rate and mean average precision (mAP) compared to traditional YOLOv7, YOLOv9, and other models, with the improved SD-YOLOv7 algorithm significantly enhancing the performance of the algorithm with minimal increase in complexity. Additionally, the model has been successfully deployed on edge devices and can successfully perform real-time detection for cooperative units. This study provides a highly practical dataset and an efficient target detection model for the field of industrial safety monitoring, and future work will explore its application in more actual industrial scenarios.