Adaptive Parameter Tuning Technique for Designing Robust Motion Detection Systems

• Main Authors: Ting-Wei Liu, 劉庭瑋
• Other Authors: Shinn-Ying Ho
• Format: Others
• Language: zh-TW
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/57314708868528927077

碩士 === 逢甲大學 === 資訊工程所 === 92 === Motion detection is one of the most important tasks in the field of computer vision. A great number of applications, from virtual reality to visual surveillance, are derived from the capability to automatically detect moving objects. The motion detection system includes motion segmentation, shadow removal and noise elimination modules. Generally speaking, systems facing pattern recognition problem rely on a large number of well-tuned parameters in order to work properly, that is, tuning such amount of parameters by manual normally requires certain domain knowledge. Current motion detection systems use a fixed set of parameters, which may influence the performance of motion detection when the environment of detected scene is changed. This paper presents an adaptive parameter tuning technique that adjusts motion detection systems automatically to deal with various kinds of environments. Also, this paper uses machine-learning technique to cluster the environments of the scene, and shows how the evolutionary approach is applied in order to automatically compute optimal parameter settings in a motion detection system to each environment group. Once the training processes are done, the motion detection system is able to automatically reason the parameter set out in order to obtain robust motion detection without any prior domain knowledge. The proposed method is demonstrated by a visual surveillance system with 7 design parameters. Experimental results are presented to demonstrate the performance and the proposed method is very suitable for practical applications.