An NFMF-DBiLSTM model for human anomaly detection system in surveillance videos

Abstract

In response to the increasing demand for an intelligent system to avoid abnormal events, many models for detecting and locating anomalous behaviors in surveillance videos have been proposed. Nevertheless, significant flaws of inadequate discriminating ability are present in the majority of these models. A novel newton form and monotonic function based deep bidirectional long short-term memory (NFMF-DBiLSTM) human anomaly recognition system was discussed in this paper to tackle those issues. Initially, videos are transformed into frames; after that, the duplicate frames are removed, and by utilizing the shannon entropy centered contrast limited adaptive histogram equalization (SE-CLAHE) algorithm, the contrast has been elevated. By using the probabilistic matrix factorization kernel density estimation (PMF-KDE) technique, the background is subtracted after estimating only the motion of the object. After this, the silhouette function is performed utilizing the dirac depth silhouette function (DDSF). In addition, clustering is done by sorting and average-based K-means (SA-KM). The features are extracted from the suspected human and are then chosen by utilizing Poisson Eurasian oystercatcher optimization (PEOO). For classifying normal or anomaly, the selected features are subjected directly into the NFMF-DBiLSTM. When contrasted with the prevailing methodologies, the proposed model is found to be more efficient.