Abstract: The anomaly detection system gives a solution to detect anomaly in crowd event video and sets alarm for public safety in mass gatherings. This paper presents a novel framework to represent video data by a set of general features, which are inferred automatically from a long video footage through a deep learning approach. Specifically, a deep neural network composed of a stack of convolutional autoencoders was used to process video frames in an unsupervised manner that captured spatial structures in the data, which, grouped together, compose the video representation. Then, this representation is fed into a stack of convolutional temporal autoencoders to learn the regular temporal patterns. Our proposed method is domain free (i.e., not related to any specific task, no domain expert required), does not require any additional human effort, and can be easily applied to different scenes. To prove the effectiveness of the proposed method we apply the method to real-world datasets and show that our method consistently outperforms similar methods while maintaining a short running time.
Keywords: Anomaly Detection; Convolutional Autoencoders, Deep Learning Technique; Convolutional Neural Network (CNN).
| DOI: 10.17148/IJARCCE.2021.101226