| 摘要: | 影像監控用於我國偵辦案件中，是一個追朔還原事實真相之重大利器，同時也適合用於追蹤過去犯案歷史軌跡，目前我國多用車牌辨識攝影機進行快速查找對象行經軌跡，然受限攝影機視角需求及成本考量等因素，無法全面汰換佈設車牌辨識攝影機;而未佈設車牌辨識攝影機之路段中，偵辦人員仍以肉眼辨識案關車輛，並調閱多個路口監視器影像，藉由外觀達成車輛追蹤之目的，因此，本論文參考車輛再識別技術，其利用車輛外觀達到不同攝影機、角度及不同時間點之影像相似度比對，達成快速過濾目標車輛之目的，應有助於快速檢索可疑車輛，並作為第一線過濾篩選。
然因車輛再識別相關研究與實務上視訊監控系統環境有些差異，主要便是標的也是以轎車為主，針對公車或是貨車之再識別需求也不高，因此，本論文為將該技術貼合實際視訊監控系統，將測試資料集先行做物件偵測，抽取車輛物件以模擬其他視訊監控系統辨識情形，並排除了公車及貨車等標記資料，形成新的資料集，再以影像連續時間相關連之車輛物件進行疊合，並成為時空相關之車身動態資料，除了利用車輛再識別之技術除了抽取訓練資料中之外觀特徵外，本論文另設計動態特徵網路層，藉由連續時間影像及車輛動態位移資料，抽取車身動態特徵，整體模型透過車身動態及車輛外觀之特徵，提升轎車類型車輛再識別之準確度，而本論文研究實驗環境係以視訊監控系統作為前提假設，對於未來使用再識別技術之研究，若也是介接至視訊監控系統，應有其參考價值。
;Video surveillance is used in the investigation and handling of cases in our country. It is a great tool to trace back and restore the truth. It is also suitable for tracking the historical trajectory of past crimes. At present, license plate recognition cameras are often used in my country to quickly find the trajectory of objects, but the camera angle of view is limited. Due to factors such as cost considerations and other factors, it is impossible to fully replace and deploy license plate recognition cameras. In road sections without license plate recognition cameras, investigators still identify suspicious vehicles with the naked eye, and call multiple intersection monitor images to identify vehicles based on their appearance. Therefore, this paper refers to the vehicle re-identification technology, which uses the appearance of the vehicle to achieve the image similarity comparison of different cameras, angles and different time points, to achieve the purpose of quickly filtering the target vehicle, which should help to quickly retrieve suspicious vehicles as the first line of filtering.
However, there are some differences between the research on vehicle re-identification and the actual environment of the video surveillance system. The main target is cars, and the demand for re-identification of buses or trucks is not high. Therefore, this paper aims to fit the technology In the actual video surveillance system, the test data set is used for object detection first, and vehicle objects are extracted to simulate the recognition situation of other video surveillance systems, and the marked data such as buses and trucks are excluded to form a new data set, which is then correlated with the continuous time of images The vehicle objects are superimposed and become the dynamic data of the vehicle body related to time and space. In addition to using the technology of vehicle re-identification to extract the appearance features in the training data, this paper also designs a dynamic feature network layer, through continuous time images and vehicles. The dynamic displacement data extracts the dynamic characteristics of the vehicle body. The overall model improves the accuracy of re-identification of sedan-type vehicles through the characteristics of vehicle body dynamics and vehicle appearance. The experimental environment of this paper is based on the premise of the video surveillance system. For future use of re-identification If the technical research is also connected to the video surveillance system, it should have its reference value. |
