Smart Video Surveillance System for Crime Prevention: Using Deep Learning

The need of existence of a better surveillance system that can detect and respond to unexpected events in real time has been needed because of the increase in criminal activity in various areas. Implementing such a system is needed because of the poor performance of the LSTM algorithm, which has low accuracy in training data. It is a need for such systems because some implementations also lacks in optimized anomaly detections. To solve this weakness a smart surveillance system is introduced in this research. The dataset that is publicly available at the University of Central Florida (UCF) includes both 20 GB of anomalous and 70 GB of non-anomalous data that can also be accessed from Kaggle. The smart system usually combines the powers of both Conv and LSTM (Conv-LSTM model) in which extracted features are fed. A function that can provide a probabilistic result like the given event is robbery, fighting, assault or more is the Softmax function. Adding to which the model achieves the ROC-AUC of 85.61% which is better compared to existing research. Providing the collaboration with deep learning capabilities and huge datasets, this smart surveillance system aims to improve precision and accuracy. This process can help in dealing with recognizing anomalous events and dealing with illegal activities around.