Generalising Violence Detection with a New Near-Real-World Violence Dataset

Detecting and classifying violence is crucial for public safety and addressing societal violence. DL models have greatly improved the automation of violence detection systems by effectively capturing intricate visual patterns. However, the quality and diversity of the training data greatly impact the effectiveness of these models. Existing datasets may be biased towards specific situations, limiting their practical use. To address this limitation, we introduce the Movie Clip (MC) dataset to enhance the generalisability of Automated Violence Detection and Classification (AVDC) systems. The MC dataset encompasses a broad spectrum of near-real-world violent actions, incorporating diverse demographics, environmental circumstances, and cultural elements extracted from movies. Consequently, it accurately reflects the complexity and diversity of real-world violent scenarios. The potential of the new dataset is investigated against its existing counterparts, like Hockey Fight (HF) and AIRTLab datasets. These datasets are used to train the ConvLSTM models based on the VGG16 and VGG19 architectures. The proposed dataset significantly improves AVDC model generalisation, outperforming the generalisability of current datasets, thereby advancing violence detection and facilitating the development of more robust and efficient AVDC systems.