Research On Key Techniques For Video Surveillance System

As the security problem gets the society more and more concerned, the video surveillance are more and more popular. The traditional surveillance systems need a lot of interactions with operators. They have little intelligent function or just very simple ones. We have reasons to believe these systems are inefficient and unreliable, because they lay too much burden on the human operators. So the objective of this research is to use what I am learning to develop some intelligent functions for surveillance systems, so that these systems can be easier to use and more reliable.The framework that I choose for the surveillance system is based on tracking. I only deal with static camera at this moment, so background modeling is a suitable way to detect the moving objects. I use Gaussian Mixture Model. Then it can decide a pixel to be foreground or background by matching its current value to the background model. Then I do the tracking job based on the detection result. It detects objects from blobs, and tracks them by matching between foreground regions and existed object models. I design a combinatorial model to get better performance. It combines an appearance template and a feature codebook. The advantages of this combinatorial model are, of course we have more tracking evidence, both a template for the whole object and many feature points as local parts. At the same time, these two models can support each other, and maintaining such a model is not expensive. We can still get real-time tracking.To analyze the tracking result, a simple but very practical method is that we define some rules. Besides, we can also try some statistical method for trajectories. I tried novel methods based on edit distance and online clustering with a sequence of Gaussian models to model a cluster of trajectories. With these methods, we can get alarm for abnormal events like loitering person and left luggage, and usual motion patterns.I have an additional work that is designed for surveillance system. That is, to find the same person in different tracks. The method that represents an object image using a bag of features has been commonly used in image retrieval and classification. In this paper, bag-of-features approach is adapted for people image description, and support vector machines are employed for high classification performance. To get more reliable matches and support supervised learning in online operation, we propose a decision scheme to distinguish never-seen individuals from reoccurrences so that the new classes can be automatically labeled. Based on these, an online recognition framework which applies incremental learning is also presented.All the proposed schemes and algorithms are tested with video sequence, and achieve their expected performances.