Crowd Localization and Anomaly Detection Using Video Anomaly Scoring Network

Abstract

Anomaly detection has become core topic in deep learning and computer vision. With the increase in surveillance cameras, it is important foridentifying the difference between the available normal data and the unusual happening in the Video Surveillance. Even though many methods are implemented by using LSTM, convolutional network and vision transformers they have high computational time, low resolution and poor anomaly accuracy.So, this work proposesa new technique named VST (Video Swin Transformer) -Anomaly Scoring Network (VASN) in high image resolution for predicting every abnormal action in each frame.University of Minnesota (UMN) dataset are pretrained in this model and then semi- supervised basedanomaly scoring network is added to find the anomaly behaviour in the video clips. The localization of crowd for target detection and then features of the target are extracted to detect the anomaly score, when comparing the other existing techniques, the work shows better performance with the score value 98.2 in the anomaly detection.