A Multi-Input Neural Network with Dense Flow and Spatio-Temporal Features for Anomaly Detection

IoT today is expanding at a rapid rate. The number of connected devices is expected to hit 70 billion by 2025 and will increase further at an exponential rate. Additionally, the surge in IoT surveillance devices makes manual human monitoring impossible and presents a pressing need for automated monitoring. Thus, we present a paper that aims to detect anomalies from surveillance videos and identify the point and duration of anomaly occurrence in the video as well. We propose a novel multi-input neural network incorporating spatio-temporal features and dense flow features to train our dataset. We show the ability of our model to capture anomalous content better than baseline approaches. The low false alarm rate shows the resistance of our model to normal videos, despite containing a lot of motion. The introduction of our new standardized dataset also opens scope for further research in this field.