| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 101 === Due to single camera finite surveillance, using amount of overlapping multi-cameras to monitor region cannot satisfy the demand of wide region video surveillance in the considerations of economic and computational aspects. In recent years, messages transformation and fusion of moving object in non-overlapping multi-cameras have become popular research in video surveillance. The difficult part of target tracking in non-overlapping multi-cameras is the spatial discontinuous of cameras, the difference of setting angles and environment of camera. Besides, people are non-rigid objects; it’s difficult for cameras to do object matching because of the external condition and the inherent psychological impact.
In this thesis, we propose an integrated system by using non-overlapping multi-cameras for different brightness and viewing angles environments to long-range tack object. The first thing is to detect the moving objects by Gaussian Mixture Model (GMM), shadow removal and morphological etc. preprocess, and then adoptive blob intersection to track moving objects. In order to deal with the objects occlusion case , we use mean shift algorithm with Kalman filter to track these moving objects. In training phase, setting up the link relation of cameras manually by the observer and using a number of known pair objects across different field of views continuously to statistics and estimate the Gaussian distribution of travel time of the objects across blind region, and further obtain the maximum/minimum travel time of the object moving through the blind region, and using cumulative BTF to get the brightness relation between different field of views. After calibrating the color of object by BTF, extract the major color of object to be the feature of object ; then combine the estimated time relation to select likely objects and match the feature of objects.
For the experiment part, we use the scenes of different illustration and view angle to analyze, such as two cameras set indoor hallway and outdoor square, three cameras set indoor hallway. The system based on the proposed method can identify objects with the accuracy of 97.5% for two cameras set indoor hallway, 94.4% set outdoor square, and 94.6% for three cameras set indoor hallway. The frame rate is about 15 to 30 fps.
|
|---|
