隨著近年來行動裝置市場的擴張，推動微型電腦系統的快速發展。車載資訊系統受益於這波風潮，在硬體方面獲得大幅提升。在此背景下，ADAS(Advanced Driver Assistance Systems，先進駕駛輔助系統)得以更廣泛的發展，已成為當前汽車電子的重要研究方向。 本篇論文關注於駕駛輔助系統中，基於車前影像進行事件警示的預防碰撞系統。當前這類研究大多嘗試對原始影像進行分析，並精確地偵測畫面中的車輛，再根據車輛與自身車的距離與狀態決定是否警示。然而，當這類方法遭遇到車輛偵測上的困難時，就可能導致警示系統失常。 本篇論文所提出的系統不依賴車輛偵測方法，而是嘗試分析單鏡頭影像的稠密光流(Dense Optical Flow)作為車前事件的特徵。我們以移動向量的直方圖統計建立不同畫面區塊的特徵向量，並利用由單純貝氏分類器(Naive Bayes Classifier)與自適應增強演算法(Adaboost)組合而成的階層式分類器(Cascade Classifier)判斷事件是否發生。此外，由於車道上的固定景物會在畫面上產生明顯的光流反應。我們也利用這個特性，將光流特徵使用於車道的偵測與分類上，以提升駕駛輔助系統對行車場景的辨識能力。 在實驗中，我們將展示本論文提出的系統對高速公路行車事件偵測具有良好的可靠度。而在車道的分類辨識上，光流特徵也確實能提供足夠的資訊，使系統能判斷自身車兩側的車道狀況。且在沒有特別最佳化的情況下，系統可以在個人電腦上達到每秒鐘30幀以上的實時運作。 ;During the past few years, Advanced Driver Assistance Systems (ADAS) are widely developed and have become an important research subject in the area of automotive electronics. In this work, we focus on vision-based collision avoidance and event warning system in ADAS. Many existing research works on this topic attempted to analyze source images and detect vehicles in the front. Then the system can make warnings depending on the distance between ego-vehicle and other vehicles. However, sometimes the vehicles cannot be detected because of the large variation of vehicle types and appearance. Also, objects which can cause dangers might not be vehicles, the system may miss these events.
This paper proposed an event warning approach based on dense optical flow analysis of monocular video rather than vehicle detection. The system constructs histograms of optical flow vectors in different regions as features. Then, cascade classifiers consist of a naive Bayes classifier and an Adaboost classifier are trained to judge events of current frame. In addition, we also attempt to use the optical flow feature in lane detection and classification and improved the abilities of driving scenario understanding.
Experiments have shown that the proposed system has high reliability of caution event detection for highway scenarios. On the lane classification part, optical flow features indeed can help the system classify lane conditions. Without specific optimizations, the system implemented on a personal computer runs at a real-time speed of 30 frames per second.
