Detecting Anomalies in Surveillance Videos Through Deep Learning and One Class Svm

Anomaly detection in surveillance videos of a crowded scenario with complex patterns is a very crucial task to ensure the safety and security of public. Supervised and Semisupervised classification approaches of deep learning to detect anomalies requires a lot of labeled data, computational units, time and it faces other limitations also. In this study, we have proposed an efficient approach of using a combination of deep feature extraction and an unsupervised classification. In the first stage pretrained Visual Geometry Group 16 (VGG16) a CNN model is used to extract deep visual features from video frames. And then in second stage One Class Support Vector Machine (OC-SVM) is used to classify these features. As OC-SVM follows the unsupervised approach it does not require labeled data and is trained on only normal frames of surveillance video. UCSD ped1 dataset is used to test the proposed method. Effectiveness of the proposed method is estimated through confusion matrix, classification accuracy and equal error rate (EER).