Detection and classification of passing objects for a security system

In this paper we compare several variations of an algorithm used to detect and classify objects passing laterally in front of a security camera. We use only a narrow and tall area of the video called the scan zone to capture the patterns of moving objects, and then we classify them by DP matching against previously stored reference patterns for all possible classes (person, bicycle, car, and bus). We test both a moving object speed independent and a speed based method for constructing patterns for the passing objects. The results show that the relatively simple method of pattern classification by DP matching can be successfully applied for classifying graphic objects of a certain degree of complexity. Finally, nonhomogencity of people's patterns and their subsequent frequent misclassification is addressed by not producing reference patterns for people, and differentiating them from correctly classified bicycles by the DP distance to the first candidate.