The paper addresses the challenging problem of perceiving meaningful activities in long video sequences, which is difficult due to the ambiguous definition of 'meaningfulness' and cluttered scenes. The authors propose two methods based on autoencoders to learn generative models for regular motion patterns (regularity) using multiple sources with limited supervision. The first method leverages conventional handcrafted spatio-temporal local features and learns a fully connected autoencoder. The second method uses a fully convolutional feed-forward autoencoder to learn both local features and classifiers in an end-to-end learning framework. The models are evaluated on various datasets, including CUHK Avenue, Subway (Enter and Exit), and UCSD Pedestrian datasets, demonstrating their ability to capture regularities and perform well in anomaly detection tasks. The contributions include showing that autoencoders effectively learn regular dynamics in long-duration videos, learning low-level motion features using a fully convolutional autoencoder, and applying the model to various applications such as temporal regularity learning, object detection in irregular motions, past and future frame prediction, and abnormal event detection.The paper addresses the challenging problem of perceiving meaningful activities in long video sequences, which is difficult due to the ambiguous definition of 'meaningfulness' and cluttered scenes. The authors propose two methods based on autoencoders to learn generative models for regular motion patterns (regularity) using multiple sources with limited supervision. The first method leverages conventional handcrafted spatio-temporal local features and learns a fully connected autoencoder. The second method uses a fully convolutional feed-forward autoencoder to learn both local features and classifiers in an end-to-end learning framework. The models are evaluated on various datasets, including CUHK Avenue, Subway (Enter and Exit), and UCSD Pedestrian datasets, demonstrating their ability to capture regularities and perform well in anomaly detection tasks. The contributions include showing that autoencoders effectively learn regular dynamics in long-duration videos, learning low-level motion features using a fully convolutional autoencoder, and applying the model to various applications such as temporal regularity learning, object detection in irregular motions, past and future frame prediction, and abnormal event detection.