異常汽車軌跡偵測以發掘徘徊事件軌跡

Abstract

由於監視攝影機與閉路電視逐漸成為公共場所的必備設施，自動車牌辨識系統於警方的應用如：偵測交通犯規、追蹤特定車輛等逐漸興盛。再加上警方朝智慧型偵辦的方針，應用自動車牌辨識系統於警方協助系統為一種可能的趨勢。 觀察利用汽車犯罪與警方目前的觀點，雖然警政署曾說明停滯、群聚與徘徊為異常汽車軌跡之常見行為，但其中最具前景的一種行為偵測──即徘徊行為，仍未在資料探勘的領域中被討論。 基於預謀犯罪事件多包含一條以上的徘徊汽車軌跡的概念，本論文首先定義徘徊軌跡的相關前提與假設，並以此為基礎發展了基於軌跡形狀的偵測機制與徘徊事件相關軌跡識別方法以偵測可能的犯罪相關徘徊行為。最後，為提升警方偵辦效率，本論文亦提供前K大可疑犯罪事件之排序準則。實驗顯示本方法的表現優於傳統異常偵測演算法DBSCAN，並具有提供事件相關資訊的優勢。

While the implementation of Automatic Number Plate Recognition (ANPR) arises due to the massive use of surveillance camera and Closed-circuit television (CCTV) in public area, numerous implementation on policing including traffic violation, vehicle tracing are adopted. In addition, since policing had evolved from crime investigation to Intelligence-Led Policing (ILP), implementing ANPR as a decision support system for policing has become a possible trend. Observing crimes through vehicle use nowadays, though National Police Agency (NPA) announced that there are several known abnormal patterns including stagnation, grouping and roam in vehicle crimes, but the most prospect part, roam behavior, are yet to be investigated in data mining. Regarding that a premeditated crime event may contains more than one roaming trajectories, we first develop a convincible premise of roam behavior through the purpose of finding premeditated crime event-related roaming trajectories, we then propose a shape-based mechanism jointed with an event-relevant trajectory identification technique to detect possible crime-related roam behaviors. Finally, Top-k possible suspicious crime events are identified and presented to aid policing. Experiment comparisons show that the proposed method performs better than DBSCAN outlier detection method. Moreover, the proposed method is advanced in providing event-relevant information.