Ladder Walking Detection via Action Recognition for Enhancing Worker Safety in Construction

The integration of AI and human-centered technologies has opened new avenues for improving workplace safety, particularly in high-risk environments such as construction sites. Ladder-related accidents, often caused by improper body posture or unsafe practices, remain a significant concern. These practices -commonly adopted by blue-collar workers attempting to save time during finishing tasks at height- are among the most frequent unsafe practices on construction sites. They significantly increase the risk of falls, and research indicates that even falls from relatively low heights are the leading cause of injury among construction workers. Such behaviors may result in falls from height, and research shows that falls from relatively low heights are the most common cause of injuries among construction workers. Due to the limited availability of real-world ladderwalking data and the challenges of safely simulating such behaviors, we curated a dataset from YouTube videos, which, while not ideal for direct safety guidance, provides valuable insights into ladder-walking techniques. Using the MMAction2 framework, we propose a comprehensive methodology that involves experimenting with multiple state-of-the-art models, selected for their effectiveness in recognizing human motion and detecting safety-critical actions in workplace environments (e.g., Temporal Segment Networks and SlowFast), optimizing hyperparameters, and evaluating performance using metrics such as accuracy. Our approach leverages action recognition to enable real-time detection of unsafe ladder behaviors, thereby reducing the risk of accidents in hazardous construction environments. This work enhances construction safety by providing a scalable, automated solution for hazard mitigation. Additionally, it highlights the broader potential of AI technologies in occupational health monitoring and risk prevention within the construction industry.