Research On Preprocessing, Target Detection And Tracking In Video Surveillance

Since the concept of Internet of Things has been proposed, people want their life moving to anintelligent direction. The research of video surveillance is critical for it is an important part ofsecurity protection. The key technology of video surveillance contains preprocessing, targetdetection and tracking and so on. This thesis will discuss the three parts above and the details are asfollows.First, we do research on the preprocessing for video image. Images are often systematicallypolluted by many kinds of noise because of the image capture, transmission and receivingequipment. Luckily, Additive Gaussian white noise and impulse noise can adequately representmost noise corrupting images. In this thesis, a new kind of filter called Weighted Bilateral Filter(WBF) is proposed to remove the mixture noise. This filter can change adaptively to removeimpulse while preserving edges and to remove Gaussian noise.Second, target detection has been studied. Target detection can be divided into two parts:moving target detection and image target detection. We can combine the two parts together. We usebackground modeling and feature extraction with classifier together to detect object, which makesthe detection more robust. In order to find a method that can extract feature from the dataautomatically, we use deep learning method to learn from the image itself without any manualwork.Third, we study the object tracking. We summarize the traditional method, and focus on theTracking-Learning-Detection (TLD) method, which has an excellent performance for long termtracking. In order to decrease the detection time of this method, we add Camshift method in thedetection procedure. Our algorithm reduces the scanning area and improves the precision, whichimproves the performance of TLD method.