主從式影像監測系統之物件追蹤演算法研究

Abstract

影像監測系統已被廣泛應用於許多場合，像是在一般公共場所或大樓出入口常見的傳統攝影機。但由於受限的監測視野範圍很小，若要增加監測視野範圍，最直接的方法就是增加攝影機的數量。但此時取得之影像會有解析度過低以致於無法用來做為辨識、判別的問題，近年來有許多研究主題為應用主從式影像系統於行人監測、通道監控，利用其可以同時提供更廣泛視野範圍的影像和擷取高解析度影像的特點。其中主控端攝影機負責監測視野範圍內的影像以及偵測、追蹤移動目標物，並且控制受控端攝影機取得目標物的高解析度影像。進行環境監測時，由於觀察的目標物並非處於靜態，目標物的動態、運動狀態、軌跡就相對重要。因此，本研究結合多顆攝影機並簡化其架構，提出針對目標物的偵測與追蹤方法，偵測移動目標物並估測其運動狀態、有效率處理多重追蹤問題並解決互相遮蔽與重疊的問題，提高系統對影像中目標物追蹤與擷取影像的準確率，其中再針對受控端所擷取的高解析度影像進行微調。本研究系統已實際應用於臺灣大學校內的戶外環境與臺北市立動物園的生態環境，並追蹤、記錄觀測場景內目標物的高解析度連續影像以及目標物移動軌跡與分佈，以提供做為離線分析的依據，其中監測視角範圍可達到195度以涵蓋監測場景之全部視野。實驗結果顯示，整體目標物的追蹤成功率約為74% ~ 77%、取得其目標物影像的解析度放大約可達到35倍，錄製之記錄影片速度可達到10 fps。

Visual surveillance system has various applications and it is an important research topic in computer vision. There are some issues of these systems on expanding field of view (FOV) and enhancing the resolution of images. Hence, the master-slave imaging systems which can provide large FOV and also high resolution images simultaneously have been applied to pedestrian surveillance, access control, and crowd statistical analysis. The master-slave imaging system is a combination of two cameras: a master camera and a slave camera. The master camera has large FOV and is responsible for monitoring and object tracking. The slave camera, pan-tilt-zoom camera, is then guided by the master camera to rotate and zoom in the targeted object to acquire high resolution images. As the targeted objects are not always stationary, the objects’ motion behavior estimation is required for tracking the correct object and acquiring zoom-in images. In reality, objects have much more complex interactions and may cause object tracking failures. In this research, data from multiple cameras in master-slave imaging system were integrated to track and predict multiple objects’ behavior. The individual object was identified and matched for occlusion handling. The multiple object tracking results show a significant improvement on the master-slave imaging system’s robustness. The system was tested at an outdoor environment in National Taiwan University campus and the Taipei Zoo. It can track and record the high resolution image sequence and trajectory of targeted object for later offline analysis. The master camera provides a large FOV of about 195 degrees. The object tracking successful rate was about 74% ~ 77% and the resolution of targeted object’s image can be zoomed to about 35 times. The video frame rate speed could achieve 10 fps.