Fight Detection in surveillance video dataset versus real time surveillance video using 3DCNN and CNN-LSTM

The abundant presence of surveillance cameras result in huge volumes of video data, which need to be monitored constantly. Real time fight detection from surveillance videos will help in preventing or stopping the fight. Fights in parking lots, bars, restaurants and public places can be avoided if there is a system that does real time detection. The proposed system compares the fight detection accuracy in surveillance video dataset by applying two approaches namely 3DCNN – Three Dimensional Convolutional Neural Network and CNN-LSTM- Long Short Term Memory network. The Surveillance Camera Fight dataset is used in this work for fight detection. 3DCNN and CNN-LSTM provide 86% and 87% accuracy respectively in classifying fight actions with the test dataset. The proposed work also includes survey and analysis of fight detection using the real time streaming surveillance video. The proposed model is tested on webcam streaming device which captures the video in real time, preprocessed and analyzed using both 3DCNN and CNN-LSTM prebuilt trained models. The real time fight detection ended up with significantly fairer accuracy with lots of challenges in implementation.