ENHANCING ANOMALY DETECTION PERFORMANCE: DEEP LEARNING MODELS EVALUATION

Detection of anomalies within video streams continues to be challenging, mostly due to the complexities involved in distinguishing abnormal activities from normal ones. This study aimed to enhance anomaly detection performance by evaluating different deep learning models and optimizers. Utilizing the Keras framework and Python on a Kaggle notebook, the experiment explored the effectiveness of DenseNet121, VGG19, ResNet50, and InceptionV3 models in conjunction with Adam, SGD, RMSprop, and Adagrad optimizers. A UCF Crimes dataset subset focused on Accuracy, F1 Score, and AUC evaluation metrics. The results establish that the InceptionV3 model paired with the Adam optimizer outperforms the other combinations, attaining AUC scores of 0.9918. In contrast to other state-of-the-art models such as DenseNet121 and ResNet50, InceptionV3 presents enhanced precision and adaptability in handling the variability found in video anomaly datasets. This study enhances security by providing insights into enhanced model-optimizer combinations, advancing video surveillance approaches, and providing support for developing robust anomaly detection systems.