Object Detection And Tracking In Intelligent Video Surveillance

Intelligent video surveillance technology is an emerging research orientation in the field of computer vision. Its main goal is to realize the analysis, understanding and description of the surveillance video sequences by computer vision technologies. Intelligent video surveillance spans many subjects including image processing, artificial intelligence, computer science, pattern recognition, and so on. The core problems of intelligent video surveillance involve the following technologies:object detection, target tracking, object recognition and behavior understanding, etc.This paper is committed to the key issues of object detection and target tracking in intelligent video surveillance systems, and some solutions are proposed for the important problems of these two technologies. The research work mainly covers the following topics:how to promote the efficiency of moving-object extraction using Gaussian mixture model, how to detect moving shadow in various scene, how to track the target robustly, and how to initialize the reference model for target tracking. The contributions of this paper can be described as follows:(1) In the procedure of moving-object extraction using Gaussian mixture model, a pixel-wised computation should be performed to obtain the parameters of the models, which lead to a high computational complexity. To promote the efficiency of moving-object extraction, an improved method was proposed for intelligent video surveillance system. Based on the fact that the majority of the background pixels remain steady, these pixels were detected according to certain criteria before the procedures of matching and updating GMM parameters, and then the frequency of matching and updating was decreased. As a result, the computation redundancy could be reduced sharply. A couple of criteria to estimate the stable and unstable status as well as the flow chart were presented to illustrate the proposed method. The efficiency is promoted significantly while keeping approximately equal quality; also it indicates the practicability of the proposed algorithm.(2) Moving shadows need careful consideration in the development of robust intelligent video surveillance systems. Moving shadow detection is critical for accurate object detection in video sequences since shadow pixels are often misclassified as object pixels, causing errors in segmentation and tracking. Aiming at Intelligent Transportation System, A moving shadow detection approach was proposed based on the change of RGB values of shaded pixels. In presented method, the coordinate systems were defined based on normalized ratio of luminance, and shadow segmentation was treated as a clustering problem in the coordinate systems mentioned above. Consequently, an ellipsoid was constructed according to scattering feature of shaded pixels. Possible shaded pixels were extracted by the computation whether they lie in above ellipsoid. Meanwhile, the relative values among the normalized luminance ratio of RGB colors were exploited to confirm shadow pixels.To increase the robustness of the shadow detection method under varying illumination conditions, an improvement is proposed for above algorithm. Illumination invariants c1c2c3 are adopted to construct ellipsoid, and then the geometrical property of shadows is exploited to perform postprocessing. These improvements provide the algorithm robustness and effect in detecting shadows for a variety of scenes and moving objects, moreover, it is suited for varying illumination conditions.(3) Exploiting color histograms and seven Hu moments, this paper presents a multi-feature tracking algorithm that adaptively weights the contribution of each feature. The weakness of the conventional color-based tracker is caused by the lack of shape information, while shape information can be described with moment values. The other advantage suggested by moments is that they can easily equipped to be invariants in 2D transformation such as translation, rotation, and scaling. Thus, the tracker can exploit the complementary information, and the combination of features improves the performance of the tracker. The proposed method is more accurate than a single-feature solution. In particular, the weakness of losing target is overcome, which improves the reliability and robustness of the tracker.(4) In this paper, we present an initialization method of reference model for target tracking. In target tracking algorithms, a reference model should be first defined in an image and then searched for in subsequent frames using a function that evaluates the similarity between the model and a candidate, In current research, the initialization of reference model is generally performed manually. The proposed algorithm can segment reference model from an image automatically and accurately, overcoming the problem in manually solution. Firstly, the targets number and position are computed using connected components labeling algorithm. Consequently, minimal circumscribed rectangles of the targets are obtained according to radius of gyration tensor method. Finally, the reference models are estimated using Monte Carlo method.