Algorithm And Application Of Moving Targets Detection In Video Streams

Detection of moving targets in video streams is one of the key research areas in digital video processing and computer vision, and also the essential technique in such primary applications as intelligent video surveillance. Among algorithms by way of background subtraction, what is based on the classical Gaussian Mixture Model(GMM) is the most popular and pervasive one, with its advantages in both facility and automatic adaptation to the dynamic background. Nevertheless, its weakness in prohibiting the emergence of any foreground moving targets during the background model initialization by way of parameters learning using sample frames, and its incapability in fast adapting to the complex changes in the background combine to egregiously constrain its application to be widely and deeply extended.It is rightly based on these shortcomings existing in the basic GMM algorithms that we present a novel framework of moving targets detection algorithm, in which we substitute the original parameters learning approach for the Local Image Flow(LIF) algorithm to initialize all the Gaussian Mixture Components for each pixel both unlimitedly and robustly, and replace the basic GMM policy for updating model components with the L-recent Window(LrW) algorithm. Apart from these two critical revisions, we also alter certain other procedures of the original GMM algorithm, including the approach to match the first model component with the largest weight, thus form the overall modified algorithm. After discussion about the algorithmic revision, we focus on the design and implementation of an applicable system of video surveillance for the marine video scenes with the video streams being transferred on the Internet, we present all that must be taken into account for an utilitarian software, and the details regarding the implementation of the novel algorithm.More emphatically, a moving shadow detection algorithm(MSDVRI) specific for applications like the video-based virtual reality interaction(VBVRI) is presented in further detail within this text. Through employing again the Gaussian Mixture to model the background, and especially introducing the two-step shadow discriminant by way of utilizing the chromatical disparity in HSV color space and difference in gray level between the pixel in the foreground and that in the background, the MSDVRI algorithm gains an appreciable reduction in the computational complexity and hence the improvement in the real-time performance of the whole detection. Besides, we modify how to judge each pixel's final category in terms of its being in background or foreground as well, so as to circumvent the time-consuming computation of variances, inclusive of correspondingly intializing, updating, and sorting all the components within the Mixture for each pixel in every incoming frame, thus improve the performance to a higher measure.Similarly, we expatiate the design and implementation of a practical VBVRI system, in which the elaborate interpretations focusing on the employment of the MSDVRI algorithm are fully presented, and an important engineering problem, i.e. the synchronization of two parallel threads that serve the shadow detection and interactive response virtualization respectively is also sufficiently probed.