Research On Object Detection Technology In Intelligent Video Survillance System

Object detection in video or images has been the focus of computer vision, it is also the key technology of intelligent video surveillance system. This study objects to target detection algorithms under the video surveillance system, in order to access a real time and accurate target detection algorithm for intelligent video surveillance system in specific environment. This article contains two modules:(1) The section of off-line machine learning-based target detection. The genery process of target detection method based on machine learning is through a large number of samples for certain types of objectives to obtain the classification for such objectives, and then use this classifier to detect the objects in images and videos. Machine learning for object detection process is more similar to human understanding and awareness of the external environment. In this section, the main research is off-line learning algorithms for target detection under certain conditions. We will introduce two features (Haar and HOG) and the the basic principles of Support Vector Machine (SVM) and Adaboost. For the sake of designing a rapid and accurate target detection system, we need to analysis the best combination of feature extraction methods and learning algorithms.(2) The part of target detection by the background modeling. In this section we briefly introduce the basic principles of Gaussian mixture model (GMM), code book (Codebook), and based on the texture based background modeling algorithm(LBP). We will analysis the advantages and disadvantages of the three classic algorithms with experiments. Combining the framework of GMM and Codebook, we design a new background model algorithm. This algorithm can estimate the probability distribution of its pixel by the sample value as accurately as GMM, while having no experience parameters and computing in real-time as the Codebook. Because the traditional LBP method is sensitive to the noise, we remove the by-bit weighted step and directly use binary LBP code to get the statistical properties of image. We design a new histogram based on Hamming distance, which can make the matching performance more stable.Finally, we summarize the full text, and propose a idea of combining off-line learning and background modeling to design an automatic target detection system. This system can reduce the workload of manually labeled samples, while increasing adaptability of video surveillance system in a wide range.