<p>Anomaly detection in surveillance videos is a critical task for ensuring public safety and enabling timely intervention. In this study, we propose an advanced deep learning–based framework for the detection and classification of abnormal activities in real-time video streams. The proposed lightweight model integrates the MobileViT for efficient classification with YOLO-v11 for accurate and fast localization of anomaly. MobileViT enables the effective features extraction on the resourse-contrained devices, ensuring the real-time classification. The YOLO-v11 accurately localized the anomalous region in the footage of surveillance. The proposed integration system provides a high performance and real time suitable solution for smart applications of the surveillance. For classification, a hybrid MobileViT model that integrates the extracted features from convolutional neural network (CNN) with vision transformers (ViTs) based on modelling of the long-range dependency. The fusion of local and global features provides better performance in the classification of anomalous events. The proposed model provides 93% accuracy for classification of the 7 different types of the anomalous events. Similarly provides 98% classification accuracy of 6 different types of the anomalous on UCF-Crime dataset, while 100% accuracy is achieved to classify the binary classes such as Robbery/normal. Following classification, the predicted anomalous frames are processed by the YOLO-v11 object detection network, optimized with task-specific hyperparameters, to precisely localize the spatial region of the anomaly within the frame. The integration of MobileViT for high-accuracy classification and YOLO-v11 for real-time localization yields a lightweight, computationally efficient, and high-performing system suitable for deployment in operational surveillance environments.</p>