Search

Understanding pedestrian flow patterns in urban areas could support the decision‐making for infrastructure planning. By incorporating computer vision techniques into surveillance video processing, human walking trajectories in a wide area could be identified by pedestrian re‐identification (ReID) across multiple cameras. Recent ReID methods mostly use convolutional neural networks equipped with deep learning techniques to extract discriminative human features from images for identity matching. However, they still suffer from realistic challenges such as occlusion and appearance variation. This paper develops a ReID‐based framework for pedestrian trajectory recognition across multiple cameras. Specifically, a generic approach of explainable model design is presented, which intuitively analyzes existing baseline models based on feature visualization. Hence, a new model named OSNet + BDB is developed that extracts discriminative‐and‐distributed features. Additionally, an incremental feature aggregation strategy is designed for more robust identity matching. Our ReID method notably outperforms its baselines by 4% identification F1 accuracy in public benchmarks. Practically, pedestrian flow statistics in a real building are extracted for behavioral modeling. Simulations of several what‐if layouts are then conducted for facility performance evaluation. [ABSTRACT FROM AUTHOR]

Construction sites are highly hazardous due to the dynamic interaction between workers and moving equipment, with high fatality rates caused by collision and falling from height, etc. Hence, identifying unsafe behaviors among workers is crucial for enhancing site safety, such as tracking their on-site movement and personal protective equipment (PPE). Vision-based video processing has been actively used to automatically recognize workers and their behaviors on construction sites. However, existing studies mainly monitor workers within a single camera capturing only a small sub-region. As workers typically move around fairly large sites, continuously tracking their movement across multiple cameras would enable more comprehensive behavioral analyses. Hence, this paper proposes a framework for monitoring safety compliance among workers, by combining worker re-identification (ReID) and PPE classification. Deep learning-based approaches are developed to address the challenges for these two tasks respectively. For ReID, a new loss function named similarity loss is designed to encourage deep learning models to learn more discriminative human features, realizing a more robust tracking of individual workers. For classifying PPE statuses, a weighted-class strategy is proposed to mitigate model bias when given imbalanced samples among classes, for improved performance despite limited training samples. By combining the ReID and PPE classification results, a workflow is developed to log any incident of workers not wearing the necessary PPEs. With an actual construction site dataset, the proposed methods improve worker ReID and PPE classification by 4% and 13% accuracies respectively, which will facilitate site video analytics and inspection of site safety compliance among workers. • Designed a similarity loss to learn discriminative human features for worker ReID • Designed a weighted-class strategy for PPE classification with imbalanced samples • Monitored worker PPE statuses with consistent identities across multiple cameras • Developed an incident logging workflow to monitor worker PPE compliance on site • Improved ReID and PPE classification by 4% and 13% accuracies over existing CNNs [ABSTRACT FROM AUTHOR]