Moving Object Detection for a Non-stationary Camera
T. Nguyen
Moving object detection with video data is a fundamental task of computer vision and image processing, which now plays an important role in many video surveillance systems. The idea is to distinguish the foreground (non-static) from the background region (static) in each frame of a video stream. Various robust algorithms such as background subtraction, optical flow, statistical approaches and temporal differencing have been introduced to address the task. However, one unresolved issue is that the background can be mistakenly considered for a moving object due to the camera movement. Therefore, in this work, the focus is on detecting and tracking moving objects in videos captured by a non-stationary camera. Although there are several proposed algorithms competing each other today, a state-of-the-art algorithm MCD5.8ms is currently the fastest one related to tracking moving objects. Unfortunately, there are two drawbacks in MCD5.8ms that need to be addressed: the relatively high amount of false positives from incorrect motion estimation, and an exceeded sensitivity to the illumination changes. In order to overcome these two challenges, in this thesis, an algorithm based on SCBU (scene conditional background update) is proposed. With the aim to boost up running time and reduce the amount number of false positives, a modification of SCBU is introduced. In addition, the implementation of SCBU will be compared with MCD5.8ms. By measuring the performance and analyzing the experimental results, it is possible to identify the advantages as well as the disadvantages of both algorithms. Moreover, potential approaches to address algorithm related issues would be elaborated on basis of these results.