基於熱影像視覺車流資料採集之電動車污染排放分析

Abstract

城市地區空氣污染水平的上升已成為許多研究的焦點，因其對人口健康的不良影響。道路上的車輛排放被確認為污染的主要來源之一。許多國家提出將車輛完全電動化作為減少道路污染的措施，然而，電動車（EVs）相較於傳統車輛對污染的實際影響仍然不確定。因此，開發一種準確區分道路上的EVs的模型可以使我們更好地了解EVs對道路污染的影響。 由於EVs和傳統車輛在外觀上沒有顯著差異，基於可見光的目標檢測方法非常不可靠。然而，熱成像技術可以準確區分這兩種車輛之間的差異。 本研究提出了一種從深度學習模型和我們收集的熱數據中進行轉移學習的方法。特別地，我們使用車輛檢測模型計算不同類型車輛的比例，並應用車輛汙染排放分析來分析EVs, EMs 對環境的汙染。 這項研究可應用於評估個人對排放物的暴露及相關健康影響。這項工作還提供了一種可靠的方法，使用熱成像技術區分道路上的EVs和傳統車輛，並可擴展到識別其他類型的EVs，如電動摩托車（EMs）、電動巴士和電動卡車。提取的數據預計還可在環境分析、交通控制、智慧城市和其他相關研究領域中提供支援。

The increasing level of air pollution in urban areas has become a focus of many studies due to its detrimental impact on the health of the population. Vehicular emissions on roads have been identified as one of the primary sources of pollution. Numerous countries have proposed the complete electrification of vehicles as a measure to reduce pollution on roads; however, the actual impact of Electric Vehicles (EVs) versus conventional vehicles on pollution remains uncertain. Therefore, developing an accurate model to distinguish EVs on roads can enable us to better understand the impact of EVs on road pollution. Since EVs and conventional vehicles have no significant visual differences, visible light-based object detection is highly unreliable. However, thermal imaging can accurately distinguish the differences among these two types of cars. This study presents a transfer learning approach from a deep learning model and an open-source dataset with the thermal data we collected. Particularly, we count the portion of different type of cars with the car detection model and applied vehicle emission analysis to analyze the emission contribution by EVs and electric motorcycles (EM). This study could be applied for assessment of personal exposure to emissions and related health impacts. This work also provides a reliable method for distinguishing between EVs and conventional vehicles on roads using thermal imaging, which can be extended to the identification of other types of EVs such as EMs, electric buses, and electric trucks. The extracted data is expected to also facilitate in different domains such as environmental analysis, traffic control, smart cities, and other related research.