Research On Object Detection And Tracking Methods In Surveillance Video

As80%of information obtained by human is from vision, it is significant to research on the intelligent system which can assist or replace human eyes. Computer vision is a subject researching on this area specifically, whose aim is to process some unhandy or effort-consuming tasks and to obtain some valuable information under specific circumstance by simulating human eyes using advanced computer technology. Intelligent video surveillance system (ISS) is an important research area of Computer vision. Two important research directions are detection and tracking of moving objects. As is fundamental and facing a lot of challenges, detection and tracking of moving objects are always the research hotspot in this area. This thesis presents two efficient background modeling methods based on deep research on detection methods of moving objects in the video sequence captured by single fixed camera. In addition, this thesis gives improvement direction of object tracking method by analyzing drawbacks of some typical tracking methods and some existing object tracking methods. Main achievements of this thesis are as follows:1. We proposed a method called Hierarchical codebook background model based on haar-like features. Firstly, the haar-like features are introduced into the background modeling area in the first time, which is used to represent a block area. Secondly, utilizing thoughts of modeling hierarchically, this model constructs two background models at block level and pixel level respectively and makes combination of these two models organically. Furthermore, we compare this model with classical background model like Gaussian Mixture Model and original Codebook model. Our approach can provide faster computation speed under the condition of same resultant effects.2. We propose modified Vibe background model using the spatial similarity between pixels. Firstly, all the operations involved in the model are integer operations. Thus it is easy to transplant to the embedded chip. Secondly, the model can be constructed using only a single frame. Initialization time is extremely short. Combined with update mechanism can result in a good solution to sudden changes of scene. In addition, the model adds shadow removal algorithm and solve the original scattered false detection problem at the same time. Thirdly, the model has great advantages in performance and computation speed.3. By researching on the classic target tracking method, we propose feasible improvement program regarding the shortcomings of these approaches.