建築工地多層次群體活動辨識

Abstract

勞動力是專案中至關重要的資源。為了確保有效的管理，管理者必須了解當前進行中的工項以及參與工人的組成情況。傳統工程中，現場建築工人生產力的分析依賴於人工取樣與記錄，不僅耗時且容易出錯。隨著電腦視覺與深度學習技術的進步，許多研究嘗試透過自動化辨識方法來解決傳統人工方法中的低效率與主觀性問題。然而，大多數現有研究僅專注於低層級的姿態辨識，忽略了建設工地協作性與動態性的特質。本研究提出了一個多層次的工班活動辨識框架，目標不僅是辨識個別工人的動作，並包括將協作的工人分組並辨識其特定的工作項目。透過運用基於圖形的表示法與自注意力機制，此框架能有效整合空間與上下文資訊，從而實現準確的辨識結果。在實驗階段，我們建立了一個涵蓋鋼筋、模板和混凝土施工作業的工地數據集，並使用多層次的指標來評估模型的性能。結果顯示，我們的框架整體 F1 分數達到 73.41%，此外，結果還顯示，即使在空間距離極為接近的情況下，模型仍能有效學習並區分不同群體。在進一步的實驗與討論中顯示，視覺特徵相似性與空間鄰近性對於準確辨識至關重要，而當這兩個要素的比重相同時，模型的性能最佳。本研究為動態施工現場監測提供了一種可擴展且高效的解決方案，同時為未來在時間建模與人-物互動分析等領域的研究奠定了基礎。

The labor force is a critical resource in construction projects. To ensure effective management, it is essential for managers to understand ongoing tasks and the composition of workers involved. Traditionally, analyzing the productivity of on-site construction workers has relied on manual sampling and recording, which is both time-consuming and prone to errors. With advancements in computer vision and deep learning, many studies have explored automated recognition methods to address the inefficiencies and subjectivity of traditional manual approaches. However, most existing studies focus on low-level pose recognition, overlooking the collaborative and dynamic nature of construction sites. This study proposes a multi-granular crew activity recognition framework aimed at not only recognizing individual workers' actions but also grouping collaborating workers and identifying their specific work items. By leveraging graph-based representations and self-attention mechanisms, the framework effectively integrates spatial and contextual information to achieve accurate recognition results. In the experimental phase, we create a construction site dataset covering rebar, formwork, and concrete operations, and used multi-level metrics to evaluate the model's performance. The results show that our framework achieves an overall F1 Score of 73.41%, moreover, the results demonstrate that the model can effectively learn to differentiate between different groups, even when their spatial proximity is extremely close. Further experiments and discussions reveal that both visual feature similarity and spatial proximity are essential for accurate recognition, with the model performing best when both factors are given equal weight. This study provides a scalable and efficient solution for dynamic construction site monitoring while laying a foundation for future research in areas such as temporal modeling and human-object interaction analysis.