Violence Detection Using DenseNet and LSTM

Detecting suspicious activities can curb increasing and varying crimes in public places by manifolds if done with accuracy. Crimes in communal areas are a global problem. Video surveillance has been in use for more than a decade but the innovative ways with which crimes are committed with every passing day, escape the human eyes. While the use of cameras for post-crime action is essential, there is a need for real-time surveillance to act as an advanced indicator to prevent or eliminate any violence before it takes place. Suspicious activities that take place in sensitive and public areas, like, transportation stations, railroad stations, air terminals, banks, shopping centers, schools and universities, parking garages, streets, and so on, with harmful intent, can be scrutinized using video surveillance and alert the nearby authorities so that the preventive measures can be taken timely. However, novel these dubious actions may be, they follow specific common patterns, including psychological oppression, robbery, mishaps, unlawful stopping, defacement, battling, chain snatching, etc. In this paper, an automated ensemble deep learning model is proposed for the recognition of possible suspicious activities. The deep learning models ensembled are DenseNet and LSTM. The model aims to train a feature extraction model for human activity recognition (HAR) for suspicious actions to achieve a high recognition rate. The research is carried out on two popular datasets for violence detection. The classification result is binary as violent or non-violent. The paper also compares the results obtained with existing methods.