A Smart Video Surveillance Technique using Artificial Intelligence to Detect Forgery and Violence in Real-Time Videos

The need for smart video surveillance has been increasing rapidly due to the ever-growing need to strengthen security and safety in complex and diverse environments. To develop robust video surveillance systems, the present paper has proposed a methodology that works in two phases. In first phase, authenticity of the video is ensured against any possible forgery attack. The proposed methodology detect the forgery by using VGG16 feature extractor on the input videos followed by cosine similarity measure. In the case, no forgery is detected in the videos, second phase of methodology works to detect any violence in video frames. In the second phase, proposed technique utilizes the Otsu thresholding-based feature extraction method along with the GFTT algorithm to identify feature points in video frames. These points are then tracked across frames using the LukasKanade optical flow technique, allowing for the capture of intricate movement patterns. The resulting clustered flow vectors are used to create a vocabulary of motion words, which is then used to map video trajectories, generating a Bag-of-Visual-Words (BoVW) representation. Finally, a Random Forest model trained on labeled data achieves high accuracy in violence detection. The proposed methodology is effective, as supported by the results, which pint to substantial improvements in the feld f surveillance technology.