Experiments and Analysis on Observation Vector Generation and Channel Number Selection in Motion Detection Algorithm Based on Independent Component Analysis

Most of the existing Independent Component Analysis (ICA) based motion detection algorithms use a single observation vector generation method and two-channel data for motion detection, which make the traditional algorithms unable to obtain a more complete and accurate state of the moving objects. In this paper, four different observation vector generation methods are proposed and larger channel numbers are introduced into traditional ICA. The motion characteristics of the moving objects are covered more widely and more information for foreground extraction is obtained from the multi-channel data. These improvements make ICA be able to deal with indistinguishable and slowly moving objects. The quantitative evaluation from different experiments shows that the multi-channel data and the combination of four observation vector generation methods enable ICA to achieve a better performance with a reasonable cost of tiny increase on false alarms.