Detecting Potential Violent Behavior Using Deep Learning

Abstract

In this master's thesis, four deep learning models - DenseNet-121, Inception-v3, ResNet50, and VGG-16 were implemented to detect potential violent behavior by applying transfer learning principles. In the theoretical part, a comprehensive review of literature in the field of human violence detection was conducted to identify prevalent strengths and gaps in existing research work. The results of the experiments conducted in this work showed the best performance with accuracy values of 98%. This work recommends, among other key findings, that future research be geared towards exploring the generalization of results from this experiment across larger datasets with adaptations to a broader domain. The future work will also benefit greatly from further hyperparameter tuning of models with more configurations.