Tools
Tools
Ardhapure, Ravindra Rajendra (2023) LSTM based Predictive Network for Video Anomaly Detection. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (700kB) | Preview |
Preview |
PDF (Configuration manual)
Download (1MB) | Preview |
Abstract
Increasing use of video surveillance has necessitated the automation of anomaly detection to provide enhanced security coverage at a lower cost due to its ubiquitous presence everywhere from large corporations to home surveillance. However, when building one such anomaly detection model, due to the absence of unexpected occurrences during training, unsupervised learning techniques have become the norm. In this study, object-level self-supervised and multitask learning is used to detect anomalous events in the video frames with the implementation of a predictive neural network framework integrated into recurrent neural networks (RNN) that learns to ascertain future data frames, allowing local predictions in each layer while only the differences between each output are passed to the corresponding network layers and this reconstruction capability can be used to make predictions. The suggested approach is flexible and can be implemented into several cutting-edge anomaly detection algorithms. To give empirical evidence of significant performance increases, the predictive coding block was integrated with long-short term memory (LSTM) RNN networks to build a prediction model that is capable of detecting anomalies and such a model can be deployed to work with any anomaly detection system in real-time with a fewer modification. The evaluation results were presented with a training accuracy of about 96.4% and 76% test accuracy and 74% test precision which asserts the model performance.
Actions (login required)
 |
View Item |