LiDAR-Based Framework for Accurate Positioning and Robust Tracking of Multiple Construction Workers

Abstract

Resource positioning and tracking are critical for progress management and safety monitoring in complex dynamic workspaces. The prevailing sensor-based approaches fall short in large-scale layout and maintenance; cameras heavily rely on light conditions and are weak in measuring distance. The latest light detection and ranging (LiDAR)-based three-dimensional (3D) tracking methods show promise in solving this challenge, but a framework specifically designed for construction scenarios is lacking. To this end, this paper presents an improved framework for tracking multiple construction workers using LiDAR based on the tracking-by-detection paradigm. The framework incorporates a worker detection module that leverages a deep learning model, featuring an improved data augmentation technique for accurate velocity estimation and a new head network to enhance the recognition of various postures. Building on the detection, the worker tracking module models complex worker movements, with a novel identity rematch strategy to maintain tracking consistency. Experiments were conducted on the LiDAR data set collected at real construction sites. The evaluation results showed that the proposed framework achieved 0.04 m positioning error and 97.4% average multiobject tracking accuracy (AMOTA) for tracking. The framework also exhibited robust tracking performance in challenging conditions such as occlusion and high crowdedness, making it a promising solution for tracking in construction scenarios.