Learning normal patterns via conv-LSTM for video anomaly detection using likelihood statistical texture feature representation in surveillance videos

An anomaly automatic detection system is a challenging issue since there is a non-deterministic assumption or definition about the abnormal events. To address this issue, this paper introduced the Likelihood Statistical Texture Feature Representation (LSTFR) method using CSR (Co-occurrence with Stationary occurrence Representation) to construct the spatial activity pattern using gray level co-occurrence matrix with likelihood estimation. Also, LSTFR is used to construct the composition histogram representation to learn the normal behaviour. The occurrence rate in a LSTFR is characterized by a histogram representation, which depends on the spatio-temporal information of the frames sequence. To efficiently classify the events using the LSTFT, this paper uses the Convolutional Long Short-Term Memory (conv-LSTM) where histogram representation of LSTFR is automatically modelled by the training with normal events. The proposed method is evaluated on four benchmark datasets: UMN, Subway, Avenue, and UCSD Ped2. The performance of LSTFR-ConvLSTM is assessed using EER and AUC-ROC, achieving superior results compared to existing anomaly detection approaches. Finally, the proposed results are compared with several existing algorithms.