Accurately detecting workers and their locations has become crucial in construction safety monitoring, but existing vision-based methods often struggle with occlusion and lack accurate spatial data, while sensor-based methods can be intrusive. This study aims to explore the potential of Light Detection and Ranging (LiDAR) for 3D construction worker detection through a comparative analysis.
First, a diverse 3D LiDAR dataset of construction sites was created. Then, several representative models were designed by revisiting and factorizing the practice of the state-of-the-art. Based on the dataset and models, a comprehensive benchmarking experiment was conducted where the candidate models were trained and evaluated under a unified framework. Their performance was analyzed using detection metrics and latency, considering the practice in construction applications.
The deep learning models achieve mean average precision (mAP) ranging from 0.856 to 0.94 for 3D worker detection. The best detection performance is achieved when applying 3D convolutional neural networks to voxel data with the anchor-free detection head. These models also demonstrate decent detection latencies at 20 ms.
The proposed method enables the 3D detection of construction workers, allowing for the continuous monitoring of spatial relationships between workers and hazardous surroundings. From a safety management perspective, this approach enriches semantic information and enhances the precision of construction management processes.
A diverse LiDAR dataset for 3D worker detection was created, which supports future 3D detection research in the construction field. Suitable design principles of 3D detection models were proposed, providing valuable foundational knowledge for application.
