Real-time crime activity detection and suspicious object identification has become very vital for public safety given the increasing demand for automated surveillance. Using the UCF-Crime dataset, this work presents an optimal hybrid model combining Convolutional Neural Networks (CNN) with AlexNet to precisely identify criminal behaviors and anomalies in surveillance films. The suggested model performs exceptionally well in spotting abnormalities by using CNN’s spatial feature extraction and AlexNet’s ability for managing vast amounts of data. Emphasizing accuracy, precision, recall, and F1-score, the Adam optimizer is used with learning rates of 0.01, 0.001, and 0.0001 to investigate the impact of training dynamics on model performance. Our studies show that lower learning rates significantly raise the model’s accuracy, obtaining on the testing dataset up to 99.53%. Capturing many real-world criminal situations, the UCF-Crime dataset offers a demanding but reasonable standard for the resilience of our algorithm. The results show that the model is fit for real-time anomaly detection in security systems since it can identify both clear and subtle suspicious activity.