Research On Algorithms To Recognize Moving Objects In Low-resolution Traffic Videos

With the fast development of computer, multi-media, and image processing technology, the techniques to recognize moving objects based on image processing have already been widely applied in intelligent surveillance systems. Compared with the traditional object recognition methods, the image processing based methods have caught widespread attention from researchers and engineers due to their advantages, e.g., no demand for additional external devices, lower overall system cost and higher scalable solution. Traffic video surveillance is an important application field of the intelligent video surveillance.Traffic videos have some prominent features, such as relatively lower resolution, complex background, various weather conditions and a variety of lighting levels. This thesis investigates the current methods regarding detection, classification and recognition of moving objects, especially the methods based on image processing technology. Meanwhile, the working conditions, as well as the advantages and disadvantages of various methods, are studied. Then, an algorithm based on multi-feature fusion and multi-frame fusion is presented for the recognition of moving objects in low-resolution traffic videos.The first step of our algorithm is background modeling and background subtraction, which aims to obtain the moving objects. For moving object extraction, some methods such as morphology processing are implemented to remove the disturbance of background noise. Then a shadow detection method based on region growth is presented to remove the most part of shadow area and to determine moving object with more accuracy. Then, it extracts some geometric features of objects from each frame image, concatenates multiple features into a feature vector and use linear support vector machine (SVM) to learn a classifier, or put these features into a cascade classifier, to yield a preliminary detection result regarding which class this object belongs to. It further fuses these preliminary detection results from multiple frames to provide a more reliable detection decision, together with a confidence level of that decision. Results in this experiments show that the algorithm based on multi-feature and multi-frame fusion is able to identify moving objects with high accuracy and low computational complexity. It is, therefore, applicable for real-time traffic video surveillance systems.Besides the recognition algorithm, the construction of our Intelligent Traffic Surveillance System is also introduced, including its main building blocks and implementation.