Deep Imbalanced Data Learning Approach for Video Anomaly Detection

Abstract

Surveillance video data often exhibit highly imbal-anced data distribution, i.e., majority or normal class instances outnumber the minority or anomalous class instances, which are the point of concern in video anomaly detection (AD). The existing deep learning methods often adopt various ensemble methods consisting of an early or late fusion of the cascade of either deep discriminative or generative learning models. These methods lack the diversity in applying the deep learning algorithms to imbalanced data learning for AD in real-world unlabeled and imbalanced surveillance video data. In this paper, decision level late fusion of two complementary deep learning models is accomplished using a loss function weighted regression model towards imbalanced data learning for video AD. Under the algorithmic level actions, the learning model's architecture consists of two complementary parallel discriminative-generative channels, i.e., a discriminative deep residual network (DRN) channel and a generative deep regression long short-term memory (LSTM) channel. The proposed complementary deep LSTM-DRN-based imbalanced data learning approach improves effectiveness in detecting anomalies compared to state-of-the-art methods.