Particle-Filter-Based Prediction for Anomaly Detection

Anomaly detection in video aims to automatically detect the unusual occurred event and model was trained from some normal samples. Previously, reconstruction-cost-based methods were the most commonly used way to solve this problem. However, because of the variety of unusual event in real world and the capacity of deep learning methods in fitting, the high cost may not always be guaranteed well on abnormal event videos. In this paper, we propose to tackle the anomaly events within a particle-filter-based method. It is based on an assumption that normal event is easy to predict, and abnormal event is hard to. We use particle filter to give the prediction’s eigenvalues and compare it with the next frame’s eigenvalues, if the prediction doesn’t conform the real data well then, we can discriminate the abnormal events from the normal data. Experiments on UMN dataset validate the effectiveness of our method.