HUMAN ACTIVITY RECOGNITION MODEL AS A VIOLENCE DETECTION SYSTEM

• Main Author: Adiwijaya Leman, Keenan
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/55948
• Institution: Institut Teknologi Bandung

In an era when many CCTVs were installed to monitor activities in public places, these CCTVs could certainly provide a sense of security because of supervision by the authorities. However, monitoring becomes difficult to do considering there are so many locations that need to be monitored continuously. To facilitate this, a system for detecting criminal activity, especially in the case of violence, could be used to assist surveillance. A spatiotemporal action localization model is a model that can be used to detect and warn authorities of violent action caught by CCTV. This research focuses on creating a model that can be the basis for the development of an efficient physical violence detection system. Efficiency becomes especially important when a model is to be applied to process many videos produced from many CCTVs. Using efficient three-dimensional convolution techniques, the use of memory resources can be minimized while maintaining decent accuracy. The model also uses multiple-scale prediction techniques that can improve the model's ability to detect small objects, such as firearms and knives. Experimental results with the UCF101-24 dataset show that the model achieves a frame-mAP score of 67.29% using only 22.84 million parameters.