Authors : Mrs. Sangita Mahendra Rajput , Prof. Dr Mangesh D Nikose
DOI : 10.46335/IJIES.2025.10.7.10
Abstract – The emergence of deep learning models has significantly improved the accuracy of anomaly detection in surveillance, crowd analysis, and other real-world applications. However, the challenge remains to balance performance with computational efficiency, particularly for real-time applications. This study assesses the performance and computational efficiency of several deep learning model architectures on the broadly used dataset: UCSD. We examine models such as CNN, Transformer, LSTM, Knowledge Distillation, Ensemble, Multi-Task Learning, and Hybrid Architectures, comparing precision, recall, F1-score, AUC, inference time, memory usage, and FLOPs. Our results indicate that while ensemble models are computationally intensive, but offer better accuracy. On the other hand, lightweight models like MobileNetV2 combined with Transformer or Knowledge Distillation maintain a balance between performance and efficiency, making them appropriate for real-time deployment. This paper provides valuable insights for selecting the right model based on the trade-offs between accuracy and computational requirements in anomaly detection tasks.
