Pedestrian Anomaly Detection Method using Autoencoder
Yingzi Wei, Xiuyun Hu · 2021
It is significant that video surveillance system is adopted for the city security. For the detection of pedestrian anomaly behavior, an autoencoder-based detection model was trained using image sequences with few or no anomalous events in the framework of unsupervised learning. Deep neural networks constructed from convolution autoencoders process video frames in an unsupervised manner for capturing the spatial information of the data. These spatial structures are then combined together to form the data representation. Three-layer convolution long short-term memory model (Long Short Term Memory, LSTM) is adopted for a temporal encoder to learn the encoding spatial structure. Finally, these feature data are scored to determine whether the pedestrian's behavior is normal according to evaluation criteria. The proposed model is trained on various datasets, which demonstrates its effectiveness and robustness.