An abnormal intrusion detection method based on self-organising model

Abstract

Intelligent video analysis is a new application direction in computer vision. Aiming at the decline of abnormal intrusion detection accuracy caused by background change in the intelligent video analysis field, an anomaly detection method for monitoring video based on a self-organising model is proposed. In this method, the idea of a self-organising mapping neural network in machine learning is applied to anomaly intrusion detection, and the gridded video image is regarded as input excitation. Then, the method constructs the expression model of a normal image by generating, updating, and deleting nodes and uses the node state in this model to calculate the anomaly degree, to judge whether there is an abnormal intrusion. Experiments show that the anomaly detection method based on a self-organising model can effectively detect the abnormal intrusion of pedestrians and vehicles in the scenes taken by a fixed lens camera and rotating ball camera. Its performance is significantly better than common detection methods. Compared with the Gaussian mixture model and grow when required network methods, the accuracy is improved by 5.8% and 2.7%, respectively.