DSLSTM: a deep convolutional encoder–decoder architecture for abnormality detection in video surveillance.

Video abnormality detection has become an essential component of surveillance video, identifying frames in the video sequences that contain events that do not conform to the expected behavior. However, their application is limited due to the presence of major challenges during training such as mode collapse, non-convergence, and instability. This paper proposes a novel two-stream spatial and temporal architecture called Deep Stacked LSTM (DSLSTM) for abnormality detection that comprises a spatial and temporal stream to extract the spatial and temporal features. MSE is computed separately for each extracted feature of the stream and fused to form the joint representation of appearance and motion. Afterwards, PSNR followed by anomaly score is measured from the joint representation. Only those frames whose anomaly score value is greater than the threshold will be considered abnormal frames. The experimental results evaluated and compared in four benchmark datasets (UCSD Ped1, Ped2, CUHK Avenue, and ShanghaiTech) depict the high performance of DSLSTM in contrast to the recent popular state-of-the-art methods. Besides, a report on three ablation experiments is also provided, and the impacts on the performance of DSLSTM are compared. We also further compared the performance of our deep DSLSTM with our own shallow SSLSTM model. [ABSTRACT FROM AUTHOR]