The Effect of Background Subtraction on Abnormal Event Detection in Video Surveillance

Abstract

In this study, the effect of background subtraction in the proposed method for abnormal event detection in video surveillance was investigated. Extracted video frames have passed through various processes and trained the model in the Long Short-Term Memory Network (LSTM) based Convolution Autoencoder architecture. There is no labeling in the data. Semisupervised learning model has been implemented by training with the normal video images. The trained model is structured according to the input data and new frames are tested on the model. The conformity of the tested frames to the model was observed with the generated regularity score. The outputs of the model trained by the background subtracted frames and the model trained by the normal frames were compared. According to the results, it was observed that background subtraction increases the accuracy in detecting abnormal events.