Faster-PPENet: Advancing Logistic Intelligence for PPE Recognition at Construction Sites

The usage of Personal Protective Equipment (PPE) is crucial in a variety of sectors, including construction, manufacturing, healthcare, and hazardous material handling, to safeguard employees from possible dangers and risks. Employers may encourage a safer workplace and lower the risk of accidents and injuries by using an automated system that can determine whether people are wearing the required PPE. Several approaches have been proposed by scientists to monitor employees at workplaces, however, these techniques are limited to recognizing a small number of PPE objects. Further, historic approaches lack the generalization power to perform well in real-world cases and are not proficient in tackling several transformational alterations of the examined images like clutter, blurring, light, and brightness changes in the background settings of images. The work is focused on overcoming the limitations of existing works by proposing an effective DL strategy called the Faster-PPENet. Clearly, we have utilized the Faster-RCNN approach with a modified ResNet101 base to detect and classify various PPE. We have presented a robust feature extractor base by altering the conventional ResNet model. Specifically, we have replaced the activation approach of the ResNet approach by utilizing the Swish activation method as an alternative to the ReLU function to extract a more effective set of sample features which are later recognized by the 2-phase locator of the Faster-RCNN model. The model is evaluated on two complex databases named the Color Helmet and Vest (CHV), and Safety Helmet Detection (SHD) to show the effectiveness of our approach. Further, a real-world data sample is also utilized to show the generalization ability of our approach. For the CHV and SHD data samples, we have attained mAP scores of 0.8845, and 0.9321, along with accuracies of 90.25%, and 94.03%, while for the real-world sample, we have attained the mAP, and accuracy values of 0.901, 91% of which is clearly proving the robustness of our approach.