Research On The Method And Application Of Video Moving Objects Extraction Based On Incremental Nonnegative Matrix Factorization

With the rapid development of technology and increase in security awareness, surveillance videos have been widely used as an important safety guarantee method. While, millions of surveillance cameras installed in public places produce huge amount of video data every day, and how to effectively analyze the video data becomes one of the researching hotspots recently. The intelligent video surveillance system can automatically analyze and understand the video data by integrated utilization of different kinds of technology, such as computer vision, pattern recognition. In intelligent video surveillance system, the accurate detection and extraction of video moving objects is very crucial. But with the increasing variety and complexity of video images, many issues and challenges appear and many applications request online method to extract moving objects. Therefore, research on real-time method of moving objects extraction in the complex scenes has theoretical significance and practical value.Another important problem in intelligent video surveillance system is how to quickly scan and index the videos. Video abstraction provides an efficient way to make the videos more compact but preserve the main information of videos at the same time. By analyzing the structure of video content, video abstraction method selects some image frames which can sufficiently generalize main content of original video data. In order to effectively store and scan surveillance video data, specialized video abstraction methods for surveillance video should be proposed.Focusing on the problems mentioned above, the major contributions of this dissertation can be summarized as follows:1. To solve the problem of extracting moving objects in complex scenes, we propose an online algorithm based on Incremental Nonnegative Matrix Factorization. Firstly in order to avoid unreasonable negative pixels appearing in the estimated background images, nonnegative constraints are added on the coefficient vector and background subspace. Besides, we introduce structured sparsity-inducing norm which cons-iders the structural relationship of object pixels to regularize moving objects. The experiments on public image database show that our method can achieve good performance for the real-time objects extraction in complex scenes.2. In order to efficiently browse the surveillance videos, a method based on moving objects extraction for surveillance video abstraction is proposed, which consists of three processing steps:video sequence segmentation, key frames extraction and movement trajectory extraction. Firstly we segment surveillance video sequence into different parts according to the moving objects on the monitor screen, and each part will be processed by the steps of key frames extraction and trajectory extraction. Then the key frames and trajectory pictures form the surveillance video abstraction. Experiments on real surveillance videos verify the feasibility and effectiveness of our method.