The large extension and complex structure of most industrial and construction areas very often make it unfeasible or inconvenient for human operators to constantly survey all the workers to detect those who do not properly wear their Personal Protection Equipment (PPE) devices. However, such a detection is of utmost importance to reduce the number of worker injuries. Consequently, the adoption of a computer vision system based on Deep Neural Networks (DNNs) that performs PPE detection by analysing the video streams from surveillance cameras is an appealing option. For instance, smart video cameras placed in the workplace might process the video frames at run-time and trigger alarms whenever they detect workers not correctly wearing PPE devices. However, in order to be sufficiently accurate, DNN-based object detection requires a high computational power that is difficult to embed in cameras. Moreover, DNN training has to be done on a large dataset with thousands of labeled image samples, and therefore the creation of a customized DNN to detect special PPE devices requires a huge effort in finding and labeling images to train the network. This paper proposes a PPE detection framework that combines DNN-based object detection with human judgement through fuzzy logic filtering. The proposed framework runs in near real-time on embedded devices and can be trained with a low number of images (i.e., few hundreds), still providing good accuracy results.

大多数工业和建筑区域的大面积延伸和复杂结构常常使操作员无法或不方便地不断调查所有工人以发现那些没有正确佩戴个人防护设备 (PPE) 设备的人。然而，这种检测对于减少工伤数量至关重要。因此，采用基于深度神经网络 (DNN) 的计算机视觉系统，通过分析来自监控摄像头的视频流来执行 PPE 检测是一个有吸引力的选择。例如，放置在工作场所的智能摄像机可能会在运行时处理视频帧，并在检测到工人没有正确佩戴 PPE 设备时触发警报。但是，为了足够准确，基于 DNN 的对象检测需要很高的计算能力，很难嵌入到相机中。此外，DNN 训练必须在具有数千个标记图像样本的大型数据集上进行，因此创建自定义 DNN 来检测特殊 PPE 设备需要在查找和标记图像以训练网络方面付出巨大努力。本文提出了一种 PPE 检测框架，通过模糊逻辑过滤将基于 DNN 的对象检测与人类判断相结合。所提出的框架在嵌入式设备上近乎实时地运行，并且可以使用少量图像（即数百个）进行训练，仍然提供良好的准确度结果。因此，创建一个定制的 DNN 来检测特殊的 PPE 设备需要付出巨大的努力来寻找和标记图像来训练网络。本文提出了一种 PPE 检测框架，通过模糊逻辑过滤将基于 DNN 的对象检测与人类判断相结合。所提出的框架在嵌入式设备上近乎实时地运行，并且可以使用少量图像（即数百张）进行训练，但仍能提供良好的准确度结果。因此，创建一个定制的 DNN 来检测特殊的 PPE 设备需要付出巨大的努力来寻找和标记图像来训练网络。本文提出了一种 PPE 检测框架，通过模糊逻辑过滤将基于 DNN 的对象检测与人类判断相结合。所提出的框架在嵌入式设备上近乎实时地运行，并且可以使用少量图像（即数百张）进行训练，但仍能提供良好的准确度结果。