Predictive Convolutional Long Short-Term Memory Network for Detecting Anomalies in Smart Surveillance.

Surveillance is the monitoring of behavior, actions, or information, with the purpose of collecting, influencing, controlling, or guiding evidence. Despite the technical traits of cutting-edge science, it is difficult to detect abnormal events in the surveillance video and requires exhaustive human efforts. Anomalous events in the video remain a challenge due to the occlusions of objects, different densities of the crowd, cluttered backgrounds & objects, and movements in complex scenes and situations. In this paper, we propose a new model called time distributed convolutional neural network long shortterm memory Spatiotemporal Autoencoder (TDSTConvLSTM), which uses a deep neural network to automatically learn video interpretation. Convolution neural network is used to extract visual features from spatial and time distributed LSTM use for sequence learning in temporal dimensions. Since most anomaly detection data sets are restricted to appearance anomalies or unusual motion. There are some anomaly detection data-sets available such as the UCSD Pedestrian dataset, CUHK Avenue, Subway entry-exit, ShanghaiTech, street scene, UCF-crime, etc. with varieties of anomaly classes. To narrow down the variations, this system can detect cyclists, bikers, skaters, cars, trucks, tempo, tractors, wheelchairs, and walkers who are walking on loan (off the road) which are visible under normal conditions and have a great impact on the safety of pedestrians. The Time distributed ConvLSTM has been trained with a normal video frame sequence belonging to these mentioned classes. The experiments are performed on the mentioned architecture and with benchmark data sets UCSD PED1, UCSD PED2, CUHK Avenue, and ShanghaiTech. The Pattern to catch anomalies from video involves the extraction of both spatial and temporal features. The growing interest in deep learning approaches to video surveillance raises concerns about the accuracy and efficiency of neural networks. The time distributed ConvLSTM model is good compared to benchmark models. [ABSTRACT FROM AUTHOR]