Research On Video Scene Classification Techniques Based On Feature Extraction

As the increasing awareness for public safety, currently the video surveillance system have been widely used. Traditional video surveillance system are performed by man-monitoring, which is not efficient, wasting valuable manpower and time. In another way, this way highly relies on the monitor's attention and subjective awareness, which makes results not absolutely correct. On this background, algorithm researching on video analysis is particularly important. Video scene classification is one of the most important video analysis algorithm, which is one fundamental but also extremely important kind of video analysis algorithm. Video scene classification algorithm offers video monitoring important reference information, hence it greatly reduces the monitoring work and increases the accuracy of monitoring results. At present this technique is mainly used to assist man-monitoring work, video data management and provide support for other video analysis techniques.Video scene classification algorithm consist of two categories: tracking-based and feature extraction-based. Video scene classification algorithm based on tracking is a traditional video scene classification algorithm. When there are too many moving objects in the same scene, this algorithm performs not so good, for the blocking and the complexity of tracking trajectory. But the video scene classification algorithm based on feature extraction is a good solution to this problem. This paper focuses on video scene classification algorithm based on feature extraction, based on the summary and analysis of current popular algorithm, finish the further researching work as follows. Firstly, this paper proposes an adaptive optical flow quantization method, which is a perfect solution to traditional optical flow quantization methods'deficiency. Optical flow vector is most popular algorithm for feature extraction, which also performs best. Traditional algorithm generally use basic and direct fixed quantization method. This method do not fully consider the distribution characteristics of optical flow, losing some information in the optical flow vectors. Based on the researching mentioned above, this paper proposes an adaptive quantization method, which chooses different partition and quantization methods according to different distribution characteristics of optical flow. As the experiment shows, this method extracts more video information from the original optical flow vectors, making the quantitative optical flow vector closer with original video characteristics and increasing the algorithm's performance.Secondly, this paper proposes a improved k-means clustering algorithm. There are two defects for traditional k-means clustering algorithm used for video scene classification, and this paper gives two solution for each one. In one hand, traditional k-means clustering algorithm randomly choose the initial clustering centers, which leads to instability in performance and efficiency of the algorithm. This paper proposes a statistics-based method to find the proper initial clustering centers. This method analyze the statistics characteristics of the original data, and it choose points that are far from each other as the initial clustering centers. This method increases the convergence speed and performance of the algorithm. On the other hand, traditional k-means clustering algorithm only consider the Euclidean distance, which usually leads to the globular clustering. To solve this problem, this paper introduces adjustable parameters in traditional k-means algorithm. This paper analyzes and finds the difference of different axis contribution for data-distances, and then define a new parameter-Euclidean distance. Clustering by using the new Euclidean distance avoids globular clustering, which is a perfect solution for traditional problems. As the experiment shows, these two methods highly increase the performance of the algorithm. Finally, this paper designs and implements the client for a telemedicine system based on ActiveX and web. This paper researches on the transferring characteristics of physiological parameters, and designs a novel transport protocol, ensuring the high accuracy and stability of the physiological parameters. Based on this, this paper designs and implements the client for physiological data receiving and displaying. This client is implemented based on ActiveX techniques, with the help of MFC framework. This paper designs a one-way voice sub-system, which is formed into a two-way voice system with H.264 system. This paper gives the ideas for design and implementation of the sub-system. Besides, this paper designs the webs for whole client system, including user management and configuration setting module. Finally, this paper shows the whole system and gives the video delay and bit rate performance experiments, which proves this system effective and feasible.