A Wrapper Approach for Video Anomaly Detection Applying Light Gradient Boosting Machine in a Multiple Instance Learning Setting

Abstract

The automatic detection of video anomalies is a challenging task due to problems such as the subjectivity of the anomaly definition and the sparseness and diversity of anomalous events with the consequent difficulty in obtaining discriminative features. Furthermore, labeling videos is a laborious and expensive task. Multiple instance learning (MIL), by labeling videos instead of frames, has become a solution to mitigate this last challenge. This work presents a wrapper-based MIL approach applying LightGBM for video anomaly detection. From the evaluation with a challenging dataset, we found that our model is competitive against other published methods which used the same test setup.