Comparison and Analysis of LSTM-Capsule Networks and 3DConv-LSTM Autoencoder in Ambient Anomaly Detection

Considering the population density in cities, manual detection of abnormal situations in the ambient by humans is an economically expensive and time-consuming application. In recent years, a number of approaches have been proposed for anomaly detection. Compared to other approaches, deep learning models have achieved great accuracy in detecting a wide variety of abnormal events. The unsupervised models outperform other deep learning models in anomaly detection. Therefore, in this study, two unsupervised deep learning models (LSTM-Capsule Networks and 3DConv-LSTMAutoencoder) are compared and analyzed with each other in the feature extraction stage. Performance is evaluated with accuracy and loss metrics. The Python programming language is used. And this work is tested with ambient anomaly datasets such as CUHK Avenue, UCSD Ped1, and UCSD Ped2. The experimental results showed that LSTM-Capsule Networks outperform the 3DConv-LSTM Autoencoder and have high potential in ambient anomaly detection among the existing methods we examined.