Violent Activity Detection on Public Transportation using Surveillance Footage
Mandar Kulkarni, Rupak Chakraborty · 2023
The increasing challenges in public management, safety, and security have made it imperative to develop an automated violence detection system. However, current AI-based techniques often face generalization problems, making violence detection difficult. Additionally, traditional training methods make it challenging for models to recognize all diverse sets of violent activities. To address this, we propose a unique approach to training models called Data Categorization. This approach involves categorizing violent actions into individual actions and training a model separately on each action subset with different learning parameters. The pre-trained models are consolidated as a network and built as a multi-model approach using ResNet 3D as the base model. The proposed method is tested in four experiments, and the results demonstrate a 2-3% improvement in accuracy compared to other state-of-the-art approaches. This study proposes a novel training method that can be applied to various data science problems and contributes to the advancement of automatic violence detection.