Abnormal Event Detection in Public Places by Deep Learning Methods

Abstract

Surveillance in public places has become an important aspect of modern lifestyle and security purposes. This is due to an increase in the number of crimes, mischievous activities and abnormal events. The monitoring process should be automated because of the excessive time consumption associated with the manual monitoring process. In general, there are many deep learning methods that exist for the classification process of abnormal events. This article provides a comparative analysis of three state-of-the-art deep learning methods used for abnormal event detection. The three methods, namely convolutional long short-term memory (CLSTM) autoencoders, convolutional autoencoders (CA) and one-class support vector machines (SVM) are tested on a benchmark dataset UCSD Ped1 which contains 34 training videos and 36 testing videos. Out of the three methods of implementation, one-class SVM offers the highest area under the curve (AUC) about 0.692 with the accuracy of 65.82%. By employing these deep learning strategies, the occurrence of abnormal events at public places like malls, roads and parks can be detected quickly, and immediate remedial measures can be implemented to enhance the public security.