Transfer learning for video anomaly detection

Abstract

Anomaly detection from crowd is a widely addressed problem in the field of computer vision. It is an essential part of video surveillance and security. In surveillance videos, very little information about anomalous behaviors is available, so it becomes difficult to identify such activities. In th is work, transfer learning technique is used to train the network. A convolutional neural network (CNN) based VGG16 pre-trained model is used to learn spatial level appearance features for anomalous and normal patterns. Two approaches are explored to detect anomalies, i) homogeneous approach and ii) hybrid approach. In homogeneous approach, pre-trained network is used to fine-tune CNN for each dataset, while testing, single dataset is considered. Whereas, in hybrid approach, pre-trained network is used to fine-tune CNN on one dataset and it is further used to fine-tune another dataset. The performance of proposed system is verified on standard benchmark datasets such as UCSD and UMN available for anomaly detection, also the results of proposed system are compared with existing deep learning approaches.

Keywords

Affiliated Institutions

• Sanskriti Samvardhan Mandal IN

Related Publications