隧道監控系統之多攝影機車輛辨識

Abstract

隧道內交通意外往往會造成巨大災害且難以處理，因此有大量監視攝影機裝設於隧道中，可即時發現事故並監控路況。但通常並沒有足夠的人力來觀看大量的監視器畫面，使得自動化監控系統的需求增加。本論文提出一種多攝影機車輛辨識系統，利用隧道內多攝影機的監視器畫面追蹤行車在隧道內的位置。 於單一監視器畫面中，使用Haar-like特徵偵測找出車輛，並取出OpponentSIFT影像特徵。接著，本論文提出的空間時間連續關係動態規劃(S2DP)演算法，利用隧道內行車順序關係性，辨識前後兩台攝影機中所偵測到的車輛。此外亦提供兩種進階辨識方法，包含即時運算(RT)方法以及非即時加強處理(OR)。即時運算方法減少車輛配對之搜尋範圍，並快速比對兩攝影機內之車輛。而非即時方法針對空間時間連續關係動態規劃演算法中無法有效配對的行車做進一步處理。 實驗結果顯示所提出之多攝影機車輛辨識系統可得到滿意的準確程度，並優於其他相關演算法。

Surveillance cameras are widely equipped in tunnels to monitor the traffic condition and traffic safety issues. Identifying vehicles from multiple cameras within a tunnel automatically is essential to analyze traffic condition through the road. This thesis proposes a multi-camera vehicle identification system for tunnel surveillance videos. Vehicles are detected using Haar-like feature detector and their image features are extracted using OpponentSIFT descriptor in single camera. The proposed Spatiotemporal Successive Dynamic Programming (S2DP) algorithm identifies vehicles from two cameras by considering the ordering constraint in the tunnel environment. Next, two methods Real-Time (RT) algorithm and Offline Refinement (OR) algorithm are proposed for different requirements. The RT fast identifies vehicles in real-time by searching a limited range of candidates, and the OR refines the identification result from the S2DP. Comprehensive experiments on various datasets demonstrate the satisfactory performance of the proposed multi-camera vehicle identification methods, which outperform state-of-the-art algorithms.