Can a machine distinguish between a high-five and a physical altercation? We brought this question to The R&D Lab to experiment on which resulted in a functional Computer Vision prototype.

Executive Summary / Abstract
This project leverages AWS SageMaker to develop and train machine learning models capable of detecting violent activity within video content. With the increasing prevalence of user-generated media and live-streamed content, there is a growing need for automated systems that can identify harmful behavior in real time. By using SageMaker’s managed infrastructure, the project streamlines the training and deployment of deep learning models at scale, enabling rapid experimentation and iteration on large video datasets.

The system processes video files by extracting key frames and motion patterns, which are then used to train convolutional and recurrent neural networks optimized for action recognition. Training is conducted on SageMaker using GPU-accelerated instances, ensuring efficient handling of high-dimensional video data. The project emphasizes not only accuracy in distinguishing violent versus non-violent scenes but also robustness across diverse environments, camera qualities, and cultural contexts, reducing the risk of bias and false positives.

Beyond technical development, this work highlights the potential applications in safety, content moderation, and security monitoring. Automated violence detection could assist online platforms in enforcing community guidelines, support public safety agencies in monitoring surveillance feeds, and provide tools for organizations that manage large-scale video archives. The use of AWS SageMaker ensures scalability and adaptability, making it possible to integrate the system into real-world pipelines while maintaining flexibility for future research and improvements.

Troy Web Consulting’s R&D Division has found that training a model to detect obvious violence (i.e. very pronounced and/or over-the-top violence) is definitely feasible, with lesser-obvious violent situations a bit more difficult to predict.

The key takeaway is that - not only is AI good at detecting broad ranges of human emotion - but that if it can accurately detect nuanced human interaction, it can likely be used to detect less-variable interactions and/or physical objects with higher degrees of certainty. Next steps for R&D is to expand the detection model to more nuanced interactions (from violence to perhaps passive aggressiveness) and then apply transfer learning methodologies to see if the model can be used for non-human (i.e. pet) interactions to detect emotion and/or violence in those videos.

Try it our for yourself at https://www.troyweb.com/lab/ai-violence-detection