使用車載攝影機之即時車輛偵測

Abstract

智慧運輸系統在近年來受到越來越多的關注。此系統的其中一個目的為保護用車人以及行人的安全。系統在進行高階的決策時，必須依賴許多較低階的資訊，駕駛附近的車輛狀態就是一個必要的輸入，其基本上構成了駕駛所面對外在環境的很大一部份。偵測車輛資訊從偵測輸入影像中的車輛區域，可以進一步應用到決定周邊車輛的位置、速度甚至於駕駛意圖。在本論文裡我們專注於單一影像中的車輛區域偵測。首先我們給出這個問題的科技發展現況，然後對最適合我們的模式：提案產生-車輛分類進行更進一步的演算法分析。對於分析中表現最好的演算法，我們實驗調整參數後對於速度及精確度的影響。最後我們提出在不同解析度下進行提案產生及車輛分類以改善執行效率的問題。

Intellectual transportation system is gaining more and more attentions in recent years. The system is designed to improve human safety for drivers, passengers and pedestrians. In order to have the ability of making high level decision the system needs to have some basic information. Detecting nearby vehicles is usually important because it provides the essential part of the driving context. Vehicle detection covers a wide range from detecting the bounding box of vehicles in the captured image to inferring the position, speed and even the intent of the vehicles detected. In this work we focus on the bounding box detection part for a single image. First we give an analyze of the state-of-the-art approaches related to this task. Then we further test some candidates of proposal generation and classification approach. And we show that the trade off between accuracy and speed can be done via adjusting parameters of these algorithms. Finally we show that by using different resolution image in proposal generation algorithm and classification algorithm can be a step toward realtime processing.