چكيده لاتين
Any unusual or rare pattern or instance in data is recognized as an anomaly. Anomaly detection has become an active area of research in recent years due to its diverse applications and challenges. Specifically, video anomaly detection has many applications, including home security, elderly care, and quality control in production lines. Our goal is to find effective methods to improve the performance of video anomaly detection.
In this research, two different methods are utilized to address the challenges of video anomaly detection. While both methods leverage a combination of different classifiers to enhance anomaly detection quality, each has its own distinct traits. The first method employs a pre-trained neural structure alongside four autoencoders that include recurrent layers in their architecture. The use of varying sampling rates allows for the analysis of frames at different time intervals, which, combined with appropriate feature extraction by the pre-trained network, creates an effective approach for anomaly detection.
The second method also combines a pre-trained feature extractor with various autoencoders for anomaly detection. However, for training these autoencoders, an adaptive adaptive boosting learning approach is used to increase the efficiency of the training process. The methodʹs architecture incorporates transformer structures and convolutional layers, which together form a robust combination for anomaly detection.
Both proposed methods have been evaluated using four publicly available anomaly detection datasets namely UCSD Ped1, UCSD Ped2, Avenue, and ShanghaiTech. The ROC-AUC scores for the datasets mentioned above are 93.5%, 95.7%, 93.4%, and 78.8%, respectively. The results of this study indicate their effective performance in anomaly detection. Additionally, through various experiments, the role of each component in the utilized structures and their performance mechanisms has been thoroughly examined.
