運用深度學習技術之道路障礙物偵測於一階段方法之效能比較

道路障礙物(Road obstacles)主要的目的在於使汽車駕駛者降低行駛速度，以保護用路人的安全，常見的道路障礙物如減速丘(Speed Bump)、減速標線(Rumble strips)等，但缺點是容易造成車輛底盤與避震器損傷。高級輔助駕駛系統隨著自動駕駛汽車的發展跟著普及，其結合攝影機、雷達、訊號發射與接收器等多種的感測器以輔助汽車駕駛員能更安全的駕駛，為日後的全自動無人駕駛汽車的到來做好準備。而高級輔助駕駛系統可同時偵測減速丘、減速標線與減速丘標誌的功能非常罕見，故本論文以此為目標建立偵測系統。本論文使用所收集的兩種不同解析度的減速丘、減速標線與減速丘標誌影像，並使用深度學習之物體偵測一階段演算法結合卷積神經網絡進行減速丘、減速標線和減速丘標誌的偵測，使用到的演算法有SSD、CenterNet、Yolov4等。每一種方法皆訓練兩種不同解析度的減速丘和減速標線影像集，直到所有演算法皆訓練結束，便進行物體偵測模型的評估，以此將最後的比對結果展現給想要自行設計輔助駕駛系的個人或組織作為參考依據，例如該使用低解析度或高解析度的攝影機作為物體偵測的輸入端，或是追求偵測的速度而犧牲一些準確率。

關鍵字

減速丘；減速標線；減速丘標誌；解析度；深度學習；物體偵測

並列摘要

The main purpose of road obstacles is to make the driver reduce the speed of the car, in order to protect the safety of the passers-by, common road obstacles such as speed bump, speed break, but the disadvantage is easy to cause vehicle chassis and shock absorber damage. With the development of autonomous vehicles, advanced driving assistance system is becoming popular. It combines cameras, radars, signal transmitters and receivers and other sensors to assist the driver to drive more safely, so as to prepare for the arrival of fully automatic driverless cars in the future. It is very rare that the advanced auxiliary driving system can simultaneously detect speed bump, speed mark and speed bump mark, so this paper establishes a detection system based on this target. In this paper, two different resolution images of speed bump, speed break and speed bump sign were collected, and the object detection one stage algorithm of deep learning combined with convolutional neural network was used to detect speed bump, speed break and speed bump sign. The algorithms used were SSD, CenterNet, Yolov4, etc. Each one way training speed bump of two different resolution and speed hump image set, until all the algorithms are training, object detection model of assessment, this will be the last of the comparison results show want designed auxiliary driving system of individuals or organizations as a reference, such as the use of low resolution and high resolution camera as the input of object detection, Or the speed of detection at the expense of accuracy.