A spatiotemporal deep learning framework for detecting unusual human activity in surveillance videos.

With smart gadgets and computers everywhere, real-time video monitoring is now necessary to keep people secure and lower the chance of unexpected human actions. However, current surveillance systems have a lot of trouble effectively spotting unusual actions in real life. Most of the advanced anomaly detection algorithms are trained and tested on synthetic or simulated video datasets. These datasets frequently do not show how complicated and different real-life settings can be. Because of this, their usefulness and capacity to be used in real life are still restricted. We propose a transfer learning-based hybrid deep framework for finding unusual activity in real-world surveillance. We test the model using raw video streams from a subset of the UCF-Crime dataset that shows actual situations of people doing things. First, a pre-trained DenseNet-201, a convolutional neural network, extracts spatial data from video frames that have already been processed. A Bidirectional Long Short-Term Memory (Bi-LSTM) network is used to describe temporal relationships between video sequences in a useful way. The spatial data and the Bi-LSTM's temporal representations are then integrated to make anomaly identification more precise and dependable. This integrated spatiotemporal architecture identifies abnormal actions of human beings in challenging monitoring situations. The model that was recommended performs better than the state of art methods. It gets 95.04% accuracy on the UCF50 dataset and 62.04% on the untrimmed UCF-Crime dataset. These results suggest that the model performs effective in real-life scenarios.