| 摘要: | 隨著交通工具使用的頻繁與流行，交通事故也日益增多。交通事故的發生，有許多情況是駕駛的分心而造成車輛偏離車道或未與前車保持適當距離而造成的。各國的交通事故統計數據顯示，前後追撞的交通事故所佔的比例相當的高；因此，我們在本研究中提出適應性車道偏離警示與夜間前車碰撞警示系統。 在適應性車道偏離警示系統中，我們計算水平第二差分影像；根據我們所定義的車道線模型，搜尋車道線，再根據車道線計算目前車輛是否偏離。而當車道線偵測失敗時，則依賴我們提出的適應性車道線偵測學習模組重新搜尋車道線，並根據連續偵測失敗的次數來調整搜尋的範圍。以這樣的方式運作，就能盡量減少運算資源來快速回復系統正常運作。最後我們將系統實作於嵌入式系統中；根據嵌入式系統的運算能力，以純整數運算實作我們的系統，讓我們的系統能在ARM 9 500 MHz的嵌入式系統上以每秒12張畫面即時運作。 在夜間前車碰撞警示系統中，我們根據車尾燈在影像中所呈現的特徵來偵測車尾燈。再以車尾燈的垂直距離、車尾燈水平距離、車尾燈移動軌跡，與兩車尾燈相關性等四項特徵配對，獲得的一組車尾燈對即視為一部車輛；最後再依據車尾燈對的距離變化量，計算與前車碰撞時間 (Time to collision, TTC)，提供駕駛人警示。在個人電腦上，運作效能可達每秒30張畫面，平均偵測正確率大約可達94.30%。 Currently, land vehicles are the most popular transportation devices in those few years, the amount of vehicles is rapidly increased, and then results in much more traffic accidents. The main factor of traffic accidents is the distraction of drivers, such as lane departure, road departure, rear collision, intersection collision, etc. Accroding to the reports of traffic accidents in many countries, the lane departure and rear collision are two important accident types; thus in this thesis we proposed an adaptive lane departure warning system and a forward collision warning system for night driving. In the adaptive lane departure warning system, we first generate a horizontal second-difference map. Then, we search the lane marks based on a pre-defined lane model in the map, Third, we judge the lane departure accroding to the detected lane marks. If the system fail to detect lane marks, the proposed adaptive learning module is launched to set proper range of line parameters for following frames. With such a learning strategy, the system will look for lane marks in the following frames as quickly as possible. In this study, we have implemented this system on a embedded system, and improved the system execution performance. The system performs an accepted frame rate of 12 frames per second with ARM 9 500MHz CPU. In nighttime forward collision warning system, we detect the tails of preceding vehicles. Then, we pair the lights using the features of: the horizontal distance, the vertical heights, the trajectory, and correlation of a pair of lights. Finally, we estimate the time to collision (TTC) of the verified light pair. The system was experimented on a personal computer and performed a high execution rate of 30 frames per second. The detection rate of preceding vehicles is also highly at 94.30% in average. |
