Designing of an Efficient Model for Violence Detection Using Advance Deep Learning Techniques

In recent times the crime rate even in the presence of CCTV is high, and surveillance footage monitored manually is increasing the chances of suspicious behaviour going unnoticed. Thus there’s a need for real-time automated analysis of footage so that such activities can be tracked and the concerned government agencies can be notified immediately, The main focus of the paper is on immediate recognition of fight scenes in surveillance videos, For this purpose three models have been tested to classify videos into fight and non-fight, ConvLSTM and 2D CNN with LSTM and 2D CNN with Bi-LSTM, The models were tested on Movie, Hockey Fights and CCTV dataset, It was observed that CNN with Bi-LSTM worked consistently well with all three datasets obtaining competitive results: 99.4% on movies, 94.8% on hockey and 73% on CCTV dataset respectively.