Urban Work Zone Detection and Sizing: A Data-Centric Training and Topology-Based Inference Approach

This study aims to address the challenges of automatically recognizing and sizing work zones in complex urban environments. We developed a deep-learning based work zone object detection model with a data-centric approach to iteratively enhance the model's performance by augmenting a custom training dataset collected from multiple sources, thereby overcoming the sparsity of annotated real-world work zone images. The training data is acquired from traffic cameras, mined from the web, and 3D-simulated work zone images. An innovative topology-based inference method is introduced, using XGBoost, for distinguishing true work zones from non-work operational zones with some work zone features. We also developed a reference-free work area size estimation method, which utilizes the standard heights of common construction equipment to provide a generalized real-pixel distance approximation. Our model's efficacy is demonstrated with an average mAP of 74.1% across all work zone classes, an accuracy of 98.4% for scene identification, and an accuracy of up to 89.52% for size estimation. Overall, our proposed approach significantly advances the capabilities of automated urban work zone detection and sizing, offering a cost-effective method to fill in the gap for the acquisition of work zone data in real-time by leveraging existing camera infrastructure.