Real-time Anomaly Detection using Tensorflow based RNN Deep Learning Classifier

Balla Mani Chidvilas, Kancharla Sai Pavan, Sanaka Naga Sai Kiran, MerugaMourya Kanth, V S Bhagavan, Tatiana Chakravorti

In this paper, a TensorFlow based Deep Learning classifier has been introduced for real-time anomaly detection to improve the commercial and national security. It becomes even more complicated and sophisticated when dealing with highly populated countries like India. Our proposed prototype consists of 4 modules namely data organization, video to frame conversion, feature extraction and model evaluation. UCF Crime Dataset is chosen as it contains a variety of video clips based on real-world anomalies to make sure the outcomes are tangible. The entire dataset of videos is transformed into individual frames using the FFmpeg library to get the features from each frame. Using Convolutional Neural Networks, the feature extraction process is made possible and the data is saved for training. The features are appended to the frames in a fixed sequence which is fed to LSTM to train the entire dataset. Added callbacks ensure that the model saves the best weights in each epoch to attain more accuracy. It has been found that the proposed model is very much acceptable for anomaly detection and gives good accuracy than some already established algorithms.

Volume 12 | Issue 2

Pages: 297-307

DOI: 10.5373/JARDCS/V12I2/ S20201046