使用軌跡分群之同調事件監控影像縮時

Abstract

監視攝影機目前被廣泛地使用來監控生活中各式各樣的事件，長時間以及大量的監視影片紀錄使得後續的瀏覽以及異常事件追蹤變得十分耗時。過去在此領域的研究為了提升瀏覽影片的效率，提出將影片中沒有事件發生的時空區塊移除並以同影片中其他事件填補，然而這樣將畫面呈現最大化的利用反而容易分散使用者的注意力。在注意力的相關文獻中顯示，當使用者必須同時追蹤影片中不同區塊的多重事件時較容易出錯，反而需要更多時間來消化影片中的資訊。因此本研究提出一種同調性的影像縮時方法來解決上述的缺點，採用事件發生的軌跡來做為同調性的依據，具有以下三個優點：(1)將相似的軌跡事件同時呈現使畫面簡潔而易於檢視與追蹤；(2)異常事件的軌跡篩選能夠簡化每個軌跡分群的內容，同時能夠分辨出可能需要警告的異常目標；(3)分群後的結果能夠有效率地搜尋特定目標。 本研究提出一套創新的分群系統，對輸入影片做背景去除演算法以及物件追蹤得到前景及其軌跡後，將所有軌跡的起點和終點做分組以得到此環境中的主要出入口位置，根據此結果配合各個軌跡的移動路徑進行分群計算，再對各個群的成員透過改良後的最長共同子序列演算法(DLCS)計算軌跡間的相似程度以篩選出群中可能存在的異常路徑。DLCS演算法保留了原始共同子序列演算法(LCS)中抗雜訊以及能夠比對伸縮軌跡的優點，同時改善了LCS演算法中無法計算兩軌跡間相異程度的缺點。得到最終的分群結果後將各個軌跡還原為影像，並計算彼此間的遮蔽率以做軌跡排程並產出壓縮影片。 本研究的實驗測試了不同場景、時間，長度的測試資料，並且在壓縮效率得到了平均73%的良好結果。在軌跡分群準確率有92%的良好結果，最後本研究提供了使用者參數調整實驗以探討不同遮蔽率下使用者個影片觀看效率比較。

To find out specific persons or objects from long surveillance videos is very labor-intensive and time-consuming. For efficiently browsing surveillance videos, video synopsis is developed to alleviate the inherent spatiotemporal redundancy by stitching objects in frames without considering time order. However, too much information in a synopsis frame may distract users’ attention. Therefore, we propose a novel surveillance video synopsis system using coherent event trajectory clustering containing the following advantages: (1) synopsis videos of similar event trajectories are concise for users to efficiently view and track; (2) abnormal event trajectory detection split out the possible suspicions; (3) the clustering results promote the efficiency of searching for a specific target. In our proposed system, events are extracted as trajectories by object detection and tracking. Starting and ending positions of all the trajectories are grouped to detect the entrance and exit regions by hierarchical clustering. Then, we cluster the trajectories by their motions of start-end positions according to entrances and exits. For trajectories in each cluster, a novel algorithm called Distance-based Longest Common Subsequence (DLCS) is designed to distinguish normal trajectories from abnormal trajectories. Finally, the trajectories in each cluster are scheduled and stitched onto the background model image to generate synopsis videos consisting of coherent event-based trajectories. Comprehensive experiments conducted on surveillance videos of various scenes demonstrate the promising results of the proposed coherent event-based video synopsis system in terms of frame reduction rate, which reaches 73% in average. And the trajectory clustering accuracy reaches 92% in average. Furthermore, the browsing efficiencies estimation under different occlusion rate thresholds are also evaluated by user subjective studies, and models of occlusion rate threshold decision are established.