One of the most prominent research areas in computer vision is video surveillance system which can monitor larger areas constantly for security purposes. Monitoring of such areas is meaningless without a prediction of abnormal situations. Human activities from surveillance videos were predicted by discovering the temporal sequential patterns. In this paper, the prediction of human abnormal activity in videos is improved by considering spatial, temporal, size and motion correlation among objects information for human abnormal activity prediction. The spatial information is collected from multiple cameras which are located at different location. The motion correlation among objects is obtained by tracking objects through particle filter technique. Before the object tracking, the objects are identified by using Modified Histogram of Gradient (MHOG). The HOG features calculate the gradient of the image using the intensity change in two directions. It will result in loss of data from the image and make the feature less discriminative because it ignores the changes of intensity in diagonal direction. But the Modified HOG features calculates the gradient of the image using the intensity change in all compass directions which make it more discriminative feature for object identification. Then, Spatio-Temporal Frequent Object Mining (STFOM) is applied to the identified and tracked objects which are encoded as a complex symbolic sequence. The frequent item sets are identified by using Frequent-Pattern tree (FP-tree) instead of Apriori algorithm. The Apriori algorithm has problems in terms of execution time and memory consumption. The FP-tree requires less execution time and less memory consumption. It returns frequent items which are considered as normal activities of human and the remaining items are considered as abnormal activities of human.
Volume 10 | 03-Special Issue
Pages: 1025-1029