Anomaly Detection in Traffic Surveillance Videos using Differential Evolution Dragonfly Algorithm with VGG16

Anomaly detection in video surveillance systems is an eminently vital field due to the widespread utility of the method, ranging from public security to monetary safety. Numerous architectures can be used for detection, extraction, selection and classification of the anomalous sequences. However, traditional optimization algorithms suffer from poor initialisation conditions and premature convergence in feature selection. Further, existing network architectures show a saturation in accuracy with an increasing number of layers. This work proposes a Differential Evolution Dragonfly algorithm (DE-DA) with VGG16 which helps to prevent local trapping for anomaly detection in traffic surveillance video. In the preprocessing step, keyframes are extracted using Histogram of Differences (HOD) to find outliers in the images and then, an auto-encoder architecture is used to perform codebook generation. DE-DA is used to select the features and VGG16 network performs classification with the help of residual blocks with skip connections. The proposed DE-DA with VGG16 achieves $31 \%$ of Equal Error Rate (EER) compared to the existing auto-encoder.