Anomalous Behavior Detection using Optimized VGG-Densenet From Crowded Video Scenes

The straight usage of anomaly detection on surveillance footage to track human activity in both private and public domains demonstrates a recent increase in research interest. The main motivation behind the installation of video surveillance systems is to secure private property and assist in the identification, prosecution, and deterrence of criminal activity, mainly in relation to terrorism in public spaces. Automated systems that can swiftly spot irregularities in crowds are essential because human-based surveillance is inefficient and more likely to make faults. Thus, an optimization-based deep learning approach for anomaly identification is designed in this work. First, videos are extracted and video frames are pre-processed using bilateral filtering. After that, object detection is done by a Convolutional Neural Network (CNN) that was trained using the Hippopotamus Optimization (HO) algorithm. Following that, the feature extraction procedure extracts features associated with color, shape and texture. Finally, VGG-DenseNet is used for anomaly detection by integration of VGG and DenseNet algorithms. Additionally, the HO algorithm is used to train VGG-DenseNet in order to improve its performance. As indicated by the experimental assessments, proposed model outperformed other traditional techniques, attaining an accuracy of 97%, sensitivity of 97%, specificity of 96%, precision of 97%, a FNR of 0.020, an FPR of 0.039, and F-measure of 97%.