Sparsity-Based Naive Bayes Approach for Anomaly Detection in Real Surveillance Videos

Anomaly detection in video surveillance data is a challenging task due to the dynamic environment of the surveillance area. Multiple solutions have been proposed to tackle this problem ranging from using handcrafted features to end-to-end deep leaning methods. This paper proposes a model-based approach to anomaly detection for surveillance video data. This solution is based on sparsity estimation of pre-trained deep features. We aim to harvest the sparsity information of C3D deep features and use it as an additional information to differentiate the normal and anomalous event in a video. The proposed approach uses the naïve Bayes model within multiple instance learning framework. Experimental results shows that using sparsity information of C3D deep features could improve naive Bayes approach and enhance the accuracy of the decision boundary.