Quantized Wavelet Dilated Convolutional Neural Network Based Intelligent Surveillance System for Public Safety in IoT Security Networks

Smart cities have produced vast amounts of surveillance video data that can be processed to identify anomalies since the introduction of visual sensors. A novel Quantized Wavelet Auxiliary Encoder Self-Attention Dilated Convolutional Neural Network with Gold Rush Optimizer (QWAESDCNN-GRO) designed to enhance security and reduce crime potential. Through the integration of the Quantized Discrete Haar Wavelet Transform (QDHWT), frame-level features have been extracted. Subsequently, the recovered features are fed to the Auxiliary Guided Knowledge Encoder-Decoder (AGED), which uses them to rank the features. These features are then sent to the Self-Attention Causal Dilated Convolutional Neural Network (SACDCNN) with Gold Rush Optimization (GRO) for anomaly detection. Performance indicators including Fl-score, recall, accuracy, precision, and AUC are compared with those of earlier techniques in order to assess the effectiveness of the recommended strategy. Using the UCF-Crime dataset, the model achieved an accuracy of 99.87% and an error rate of 0.021%. On the Hockey Fights dataset, it obtained an accuracy of 99.88% and an error rate of 0.019%.