目前國內高速公路各路段均設有 CCTV，這些監視器畫面均透過網路方式開放予民眾使用，但對於這些監視畫面僅能透過人為觀察來判斷是否塞車或是發生事故。在過去的研究裡已能利用這些監視畫面，透過深度學習裡知名的 YOLOv3 演算法來辨識監視畫面內之車輛，期望透過辨識結果來偵測或是預測該路段是否塞車。但由於當時使用於辨識的模型是由 COCO (Common Objects in Context)資料集訓練而成，且因為高速公路監視畫面的解析度並不高，因此研究結果對於監視畫面內的車輛辨識程度不如預期，亦會有誤判的情況發生。因此本研究為解決原先YOLOv3不足的缺點，欲使用YOLOv4來改善其功能，其方式為將原本辨識的影像結果儲存下來，經人工判斷排除誤判結果影像後，來建立更適合於高速公路監視畫面的車輛影像訓練資料集。再搭配深度學習之影像增強技術將監視畫面放大，以及影像辨識方法裡的YOLOv4演算法，配合參數調整來訓練自己的影像辨識模型，以成為更適合於辨識高速公路監視畫面內車輛的深度學習模型。經實驗結果顯示，本研究重新訓練後的 YOLOv4 模型，對於 car 類別物件偵測辨識率為 93.8%，bus 類別物件偵測辨識率為 65.1%，而 truck 類別的物件偵測辨識率為 40.82%，整體平均錯誤率由 18%降低至 14%。而 car 類別 F1 數據由訓練前 0.93 提升至 0.96；bus類別 F1 數據由訓練前 0.63 提升至 0.8；truck 類別 F1 由訓練前 0.58 提升為 0.62，表示本研究所提構想經實驗證實可行，表示本研究所提構想經實驗證實可行。 Currently, CCTV is installed on all sections of domestic highways, and these monitors are available to the public through the Internet. However, these surveillance images can only be observed by humans to determine whether there has been a traffic jam or an accident.Previous studies have been able to use these surveillance images to identify vehicles in CCTV images using the well-known YOLOv3 algorithm in deep learning, with the expectation of detecting or predicting traffic jams on roadways through recognition results. However, since the model used at that time was trained from the COCO (Common Objects in Context) dataset, and the resolution of the highway CCTV images is not high, the results of the study were not as good as expected for the recognition of vehicles in the CCTV images, and there were some false positives.Therefore, this study saves the original image recognition results and builds a more suitable vehicle image-training dataset for highway CCTV images after eliminating any misjudged images through manual judgment. We then use the deep learning image enhancement technique to enlarge the CCTV images and the YOLOv4 algorithm in the image recognition method to train our own image recognition model with parameter tuning to become a deep learning model more suitable for recognizing vehicles in highway CCTV images.Experimental results showed that the retrained YOLOv4 model has a 93.8% object detection recognition rate for the car class, 65.1% for the bus class, and 40.82% for the truck class, while the overall average error rate has been reduced from 18% to 14%. And the F1 number of car class improved from 0.93 to 0.96; the F1 number of bus class improved from 0.63 to 0.8; the F1 number of truck class improved from 0.58 to 0.62, indicating that the proposed concept has been experimentally proven to be feasible
