Object Identification In The Situation Of Multiple Cameras

With extensive use of surveillance cameras, a large amount of information in surveillance videos is required to analyze and process by manpower or machine. Currently, an urgent requirement for the intelligent video surveillance system with multiple cameras is made in the security field. Meanwhile, the object identification and tracking techniques with multiple cameras become one of the most important directions for domestic and foreign researchers engaged in the video tracking field. This thesis studies the object identification approach in the situation of multiple cameras, which is the core technique in this field and determines the performance of multiple cameras surveillance systems. The thesis focuses on the identification techniques, including the object identification of combining multiple cameras’color and texture features, the object identification based on the spatio-temporal relationship and the object identification integrating the color, texture and spatio-temporal information under the Bayesian framework. The main work and innovation of this thesis are shown as follows:(1)The texture feature applicable to multiple cameras is proposed, which is based on the SIFT matched point set and the extension of2bitBP characteristics. The texture feature (appearance information) algorithm present in this thesis is able to adapt to changes in the multiple cameras’ attitudes and angles. A certain degree of differentiation towards different objects with similar color distribution is obtained and this algorithm has satisfactory reliability.(2)Object identification algorithm combining multiple cameras’ color and texture features is studied. For the identification of color feature, a color histogram matching algorithm on the basis of brightness transfer function is applied and the matching degree among targets can be calculated using Bhattacharyya Distance.(3)The topology structure of multiple cameras with both overlapping and non-overlapping fields of view is studied and represented by matrix. Relational matrix is used in the description of the relationship of the multiple cameras marked manually, which can deduce the topology structure of multiple cameras quickly and accurately. This approach guarantees the reliability of spatio-temporal relationship learning.(4)Aimed at the surveillance network of multiple cameras with both overlapping and non-overlapping fields of view, a unified identification approach based on spatio-temporal relationship is described in this thesis. The traditional transfer time probability distribution is used to express the spatio-temporal relationship of the multi-camera with non-overlapping fields of view. An improvement on the traditional transfer time probability distribution is made in this thesis. The negative value is introduced into transfer time, which makes traditional transfer time probability distributed in the real space and solves the spatio-temporal relationship of multiple cameras in the case of both overlapping and non-overlapping fields of view.(5)The color, texture and spatio-temporal information of the object are integrated to realize the object identification system of multiple cameras. In addition, the system achieves the function of object identification and tracking with good reliability and high accuracy, which demonstrate the effectiveness of the method proposed in this thesis.