PERFORMANCE EVALUATION OF ANOMALY DETECTION MODELS USING LPQ ENSEMBLE SP-HOG, STAP, AND STATISTICAL FEATURES ACROSS VARIOUS SURVEILLANCE DATASETS

This study presents a comprehensive evaluation of anomaly detection models utilizing advanced feature extraction techniques, including Local Phase Quantization Ensemble Spatial Pyramid Histogram of Oriented Gradients (LPQ ensemble SP-HOG), Space-Time Adaptive Processing (STAP), and Statistical Features. The models were applied to diverse surveillance datasets such as UCSD Ped2, CUHK Avenue, Shanghai Tech Campus, UCF-Crime, and Street Scene. Performance metrics, including accuracy, precision, recall, F1 score, AUC, log loss, specificity, balanced accuracy, and cross-validation accuracy, were analysed. Results indicate that Optimized Deep Convolutional Neural Networks (CNNs) consistently outperform traditional machine learning and standard deep learning models, setting a new benchmark in anomaly detection for surveillance videos. The findings underscore the importance of advanced feature extraction techniques and optimized deep learning architectures in enhancing anomaly detection capabilities.