<p>Video surveillance systems play a pivotal role in enhancing security and traffic management in smart cities. Detecting and localizing anomalies in video data remains a challenging task, especially given the need for real-time processing and high accuracy. In this paper, we propose a novel framework that leverages a dual-stage attention mechanism for robust video anomaly detection and localization. Our approach integrates Convolutional LSTM (CLSTM) with the Convolutional Block Attention Module (CBAM) to predict future frames in video sequences. The proposed Dual Stage Attention Prediction Module (DSAPM) enhances the model’s ability to focus on critical spatial and temporal features, significantly improving anomaly detection performance. We utilize the prediction error between the predicted and original frames as a decision parameter, optimized by a Radial Basis Function Network (RBFN) to ensure consistent and accurate frame-level anomaly detection. To refine anomaly localization, we introduce the Dual Attention SpyNet (DAS), which combines SpyNet with CBAM to analyze motion patterns and localize anomalies at the pixel level. Experimental results on benchmark datasets, including UCSD Ped1, UCSD Ped2, Avenue, and ShanghaiTech, demonstrate the effectiveness of our approach, achieving AUC scores of 98.05%, 98.21%, 97.00%, and 96.42%, respectively. Our framework provides a robust solution for unsupervised anomaly detection in complex video surveillance scenarios. Link to the code: <a href="https://github.com/DevashreePatrikar/DSAPM_RBFN_DAS.git">https://github.com/DevashreePatrikar/DSAPM_RBFN_DAS.git</a></p>