破碎的地質構造、頻繁發生的地震，再加上雨季與颱風季充沛的降雨，使台灣公路邊坡的穩定度受到極大 威脅。山區公路容易發生邊坡災害，對國家經濟與交通命脈造成危害。一般公路維護作業大多仰賴巡路人 員，以步行或乘車的方式，目視檢查道路或邊坡狀況，但這種傳統方式不但耗費大量人力與時間，也無法 針對路況即時提供相關的空間資訊。除了缺乏客觀量化的科學數據與舊有資料進行比對，也難以偵測到邊 坡災害細微但持續性的前兆。本研究依序分析了2011年4月20日與2011年11月22日於中部橫貫公路拍攝的環景攝影影片，選取出12處易發生邊坡災害的地區，再從資料中擷取出每個地區的多期環景影像進行變異分析。由於一般的巡路車輛並未收錄精確的GPS和IMU資訊，上述兩段環景影片也不是從相同視角拍攝，因此本研究整合了三種幾何配準方法處理多期環景攝影資料。影像先經過自適化增揚處理，再使用影像匹配技術SIFT方法萃取特徵點，然後利用交互相關方法（CC)與相位相關方法（PC)檢驗特徵點並濾除掉有問題的特徵點。留下可用的特徵點再以PC產生匹配點，每一點都以CC檢驗，精確的匹配點讓多期環景影像能精確的配準以進行變異分析。研究結果顯示多期環景影像在配準後所偵測到的變異處，能針對公路邊坡的細微變化提供可靠的量化資訊，此項自動化的變異分析技術為公路邊坡提供了創新且費用低廉的安全評估方式。

The broken terrain and frequent earthquakes, together with the heavy precipitation during the rainy and typhoon seasons, pose a grave threat to slope stability in Taiwan. As a result, slope disasters are frequently found along the highways in mountainous area and seriously endanger Taiwan’s lifeline of transportation and economy. The traditional approach for highway maintenance relies on patrolmen to visually screening the slopes from the ground or the patrol vehicle. Such an approach, however, requires considerable manpower and time, yet provides very limited information on spatial coverage. Lacking of an objective and quantitative comparison between the latest observations to the historical one, there is no way to diagnose the subtle yet progressive signs of slope disasters. This research employs two panorama videos of New Central Cross-Island Highway, taken on 20 April 2011 and 22 November 2011, respectively. A total of 14 sites with high risk of slope disasters are identified and selected. The multi-temporal panoramas of each site are extracted from the videos for change detection. Since the accurate GPS and IMU data were not recorded in an ordinary petrol vehicle, and these two videos were not taken from the same viewing angles along the same route, we integrate three approaches to coregister the multi-temporal panoramas. First, the adaptive enhancement is applied to the multi-temporal panoramas and scale invariant feature transform (SIFT) approach is used to generate a set of key points. These key points are examined by both the crosscorrelation (CC) approach and the phase-correlation (PC) approach, with the intention to fill out those problematic points. Based on these robust key points, the PC approach is used again to generate a large number of tie points and each point is double checked with CC approach. With the large amount of accurate tie points, the multi-temporal panoramas can be accurately coregistered to meet the requirements of change detection. The results demonstrate that the difference between the coregistered multi-temporal panoramas provides reliable and quantitative information of subtle changes on highway slopes. This processing can be carried out in a fully automatic fashion, which is an innovative and low-cost approach to assess the safety of highway slopes.