The Worker Risky Behavior Detection in Industrial Scenarios Based on Deep Learning

In industrial environments, ensuring the safety of workers is a core concern for both enterprises and society. Incidents resulting from hazardous behaviors of workers lead to severe impacts on both society and families. Therefore, it is particularly important to develop accurate and efficient methods for detecting workers' hazardous behaviors. Currently, detecting hazardous behaviors of workers based on video surveillance data is the primary method. However, traditional manual monitoring is time-consuming and laborintensive, making it difficult to ensure the accuracy and timeliness of detection. With the development of computer vision technology, existing deep learning methods have been applied to such tasks due to their ability to automatically extract features. However, through field surveys and extensive research, it has been found that existing studies typically construct datasets based on specific application scenarios, which often focus on a single type of hazardous behavior and cannot meet the demands of complex industrial environments. Furthermore, various types of hazardous behaviors occur rapidly in real scenarios and face challenges such as complex backgrounds and small targets. Therefore, ensuring detection accuracy while considering timeliness has become a key issue that needs to be urgently addressed. In response to these key issues, this thesis focuses on safety helmet detection, with the aim of improving detection timeliness and accuracy through research.