Motion analysis of temporal features in video surveillance

• Main Author: Yuan, Kirsten Shaoqing.
• Other Authors: Ang Yew Hock
• Format: Final Year Project
• Language: English
• Published: 2009
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: http://hdl.handle.net/10356/16780
• Institution: Nanyang Technological University

Ever since the various terrorism attacks, enforcing security have become a world wide issue. As Closed-circuit television (CCTV) enables surveillance of multiple areas in real time without physically being there, thus, more CCTV is being installed in banks, schools, train stations, corporations, shopping centres and government offices. However, this means that we need more man power to surveillance over the large amount of streamed video which could be costly and time consuming. With limited human concentration span, it is impossible to notice all possible threats and crimes. In this project, it aims to auto make the process of video surveillance system through precise analysis of the images captured. In this report, the theoretical aspects of computer vision techniques used in the development of the prototype are explained in details. These techniques include pre-processing techniques such as gryscaling and median filtering, image differencing, edge detection methods and automatic thresholding. A review of existing computer vision systems that are based on image differencing, feature recognition and model based recognition are discussed and evaluated. Some of the techniques used in these systems were also explored in the development of the proposed system. The technical aspects of the prototype are also discussed and evaluated. The proposed system is able to read both video or picture files and process them to obtain analysis on detection. The system is built primarily on image differencing methods and it is able to recognise the region of motion, the foreground objects as well as provide a measure to classify the congestion level by computing the pixel area ratio. Stationary vehicles were also detected using this method. The evaluation of the system showed that the proposed system is able to reflect a distinct difference between normal traffic flow and heavy traffic flow. However, the accuracy of the system is subject to weather and lighting conditions of the scene. Also, the speed of the current system is not ideal for real time application. Future recommendation includes development in shadow removal functions.