Towards database-free vision-based monitoring on construction sites: A deep active learning approach

In order to achieve database-free (DB-free) vision-based monitoring on construction sites, this paper proposes a deep active learning approach that automatically evaluates the uncertainty of unlabeled training data, selects the most meaningful-to-learn instances, and eventually trains a deep learning model with the selected data. The proposed approach thus involves three sequential processes: (1) uncertainty evaluation of unlabeled data, (2) training data sampling and user-interactive labeling, and (3) model design and training. Two experiments were performed to validate the proposed method and confirm the positive effects of active learning: one experiment with active learning and the other without active learning (i.e., with random learning). In the experiments, the research team used a total of 17,000 images collected from actual construction sites. To achieve 80% mean Average Precision (mAP) for construction object detection, the random learning method required 720 training images, whereas only 180 images were sufficient when exploiting active learning. Moreover, the active learning could build a deep learning model with the mAP of 93.0%, while that of the random learning approach was limited to 89.1%. These results demonstrate the potential of the proposed method and highlight the considerable positive impacts of uncertainty-based data sampling on the model's performance. This research can improve the practicality of vision-based monitoring on construction sites, and the findings of this study can provide valuable insights and new research directions for construction researchers.