This paper describes a method of detecting intruders by video image processing. There are some studies on detecting intruders in video image by analizing the difference-binary images and motion vectors. These methods are, because of using simple features such as size and motion vectors of changing regions, not satisfactory to reduce erroneous detections caused by other factors, such as small animals, sways of trees, and changing in brightness.
We propose a method to detect intuders by time series data by projection pattern of silhouette. In this method, changing regions in video images are extracted by interframe differential, and are transformed into projection patterns. Evaluation scores (a measure of similarity to human's silhouette) are caluculated by flexible matching between observed patterns and the reference pattern. Recognition processing is carried out using time series score by tracking a target of each video frame.
We have made experiments with a prototype and confirmed the effectiveness of this method.