Human Crime Anomaly Detection using Deep Learning

The research focuses on developing an automated Human crime action recognition system using hybrid deep learning models to classify various anomalous human crime activities from real-time videos. We have used UCF-Crime dataset that contains 13 classes of anomalous activities and normal videos class, the research work explores and compares the performances of various models like Resnet-50 combined with simple-RNN, GRU, LSTM with models from other similar studies. These models were trained for binary, 5 -class and 14 class classifications, outperforming others in all the three respective classes. Thus, the research demonstrates the potential of deep learning in enhancing surveillance security by automating the detection of various serious crime actions in real time, offering significant improvements over traditional methods of manual surveillance.