Integrating clustering with LSTM diffusion for video anomaly detection

Automatic anomaly detection is very important in smart video surveillance because it helps identify unusual events or actions that do not match normal behavior. This study presents a new approach called Color Histo-Sigmoid Fuzzy C-Means (CHSFCM) as the main idea to improve the accuracy of object separation and background modeling. The method has three steps: first, it uses CHSFCM for pre-processing and segmentation to better understand the background. Second, it uses a Deep Neural Network (DNN) to extract features. Third, it uses a Support-based Long Short-Term Memory (SLSTM) network to tell the difference between normal and unusual events. By combining the strong ability to separate different parts of data from CHSFCM with the time-learning power of SLSTM, the system performs very well, achieving an accuracy of 96.25 %, which is better than other methods. This approach gives a reliable and strong way to spot unusual events in real-time video monitoring. • This study found that SLSTM classification, DNN-based feature extraction, and CHSFCM clustering are the best DL methods. • This research presents a Support-based LSTM network that detects normal or abnormal video event in temporal dynamics. • The proposed method enhances current methods and is effective in practical applications with 96.25% accuracy. • After validating the HARD-ATM dataset, the vast UCF-Crime dataset was used for robustness and generalizability.