使用移動向量分析之車用來車偵測與碰撞警示系統

Abstract

為了降低在日常生活中交通事故所造成的傷亡，藉由雷達或相機的車用來車偵測與碰撞警示系統已在近年逐步受到重視。本論文提出一套適用於日間的車輛偵測系統，該系統利用架設在車輛後方是數位相機擷取路面的影像資料，並分析畫面中各方塊之移動向量場，分割出畫面中屬於路面和屬於來車的部份，並藉由估算來車與駕駛車輛的距離，依照距離長短不同而給予不同程度的警報。 然而，利用傳統的方式估算移動向量場時，會因為投影變換的誤差以及路面材質強度的不足，移動向量場無法被準確的估算。本論文提出一套演算法，可在不影響來車部份的移動向量場的狀況下降低路面部份移動向量場的誤差。藉由移動向量場分割的結果，我們可以進一步的和色彩分析的結果整合以估計來車的距離和方位，最後可在大約二十公尺的範圍內有效估計來車的存在。

On road vehicle detection and collision warning system has attracted more attention these years in order to minimize the traffic accidents in the daily life. A vehicle detection system applicable in daytime is proposed in the thesis. In this system, a digital camera mounted behind the vehicle to capture the image sequence. Our purposed system then analyzes the block based motion vector field, segment out the vehicle region from the road bed region, and then estimates the relative position between the driver and the vehicle. The system is also able to assist the driver by sending out a warning signal according to the estimated relative distance. However, tradition block based motion analysis does not work properly due to the perspective distortion and low texture on the road area. We purposed an algorithm to rectify the motion on the road plane without affecting the motion accuracy on the vehicle region. Segmentation is done by analyzing the motion vector. To increase the accuracy of the results, color based segmentation is also implemented to fuse with the motion based segmentation, and the experimental results showed that the proposed system is able to detect the existence of vehicles effectively in a distance of about 20 meters.