基於影像辨識與氣象資料之熱舒適度衣著隔熱值模型

Abstract

本研究旨在探討多角度衣著辨識對衣著隔熱值估算之影響，並建立一套結合氣象資料的調整模型，以提升熱舒適度評估的準確性。熱舒適度為反映室內環境品質的重要指標，而衣著隔熱值是熱舒適度評估模型中的個人參數之一。現行研究常依據季節設定固定衣著隔熱值，然而實際穿著狀況會受到文化背景、個人習慣與氣候條件等影響，導致評估結果與實際情形產生落差。 本研究使用 YOLOv7 模型，並以 DeepFashion2 資料集進行訓練，以辨識 13 種常見服裝類型並估算衣著隔熱值。實驗設計中，受試者進行 360 度旋轉，由四部固定攝影機於不同高度與角度同步拍攝影像，分析不同視角對辨識結果的影響。結果顯示，側面與高角度影像中辨識誤差較高，顯示視角變化對衣著隔熱值估算具有重要影響。此外，模型僅能辨識可見的外層衣物，無法辨識內層衣物，導致多層穿著普遍低估衣著隔熱值。 為修正上述偏差，本研究建立了一套衣著隔熱值調整模型，整合天氣資料與辨識結果，透過深度神經網路進行修正預測。模型在交叉驗證中展現穩定預測能力，並於實際監視器安裝位置所拍攝之正面影像進行初步測試，顯示其具備應用於真實場域的可行性，為未來智慧建築中的熱舒適度即時管理提供參考。

This study explores the impact of multi-view clothing recognition on clothing insulation estimation and develops an adjustment model that integrates meteorological data to enhance prediction accuracy. Thermal comfort is an important indicator reflecting indoor environmental quality, and clothing insulation is one of the personal parameters in thermal comfort models. Existing studies often assign fixed clothing insulation based on seasons; however, actual clothing conditions are influenced by cultural background, personal habits, and climate, which may result in differences between the estimated and actual conditions. The YOLOv7 model was trained using the DeepFashion2 dataset to recognize 13 common clothing types and estimate clothing insulation. During the experiment, participants performed a 360° rotation while being photographed by four fixed cameras positioned at different heights and angles. The results show that recognition errors are notably higher in side and high-angle views, indicating that viewing angle plays a significant role in clothing insulation estimation. In addition, the model can only detect visible outer garments and is unable to identify inner layers, which often leads to underestimation of insulation values in multilayer clothing scenarios. To address these limitations, a deep neural network adjustment model was developed by integrating meteorological data and recognition results. The model demonstrated stable performance through cross-validation and was tested using front-view images captured at typical surveillance camera positions, showing its potential for real-world application. This study offers a practical reference for real-time thermal comfort management in smart buildings.