In the digital world, children are increasingly exposed to online video content, including materials that may be violent and harmful. Existing approaches, such as object detection and sequential models, often struggle to capture both spatial and temporal f eatures in video. This limitation frequently results in reduced accuracy, especially when analyzing complex or fast – paced scenes. To address this gap, this study developed WebVidGuard, a real – time video violence detection system in video streams. WebVidGua rd combined a 3D Convolutional Neural Network (3D – CNN) with a Gaussian Mixture Variational Autoencoder (GMVAE) to improve detection performance. WebVidGuard was a web – based application that automatically detected and blocked violent video content during pl ayback in real time. The study adopted a mixed – methods approach, combining experiments and development. A dataset of 2,000 video clips, with 1,000 violent and 1,000 non – violent videos, categorized into Punching, Kicking, Head – Hitting, Shooting, and Normal Videos Classes. The d ataset was split into 60% training, 20% testing, and 20% validation sets. The system was trained and evaluated using the 3D – CNN and GMVAE models, and performance was measured using accuracy, precision, recall, and F1 -score using a confusion matrix analysis across five (5) evaluation runs. The results show that WebVidGuard attained high accuracy and efficiency in detecting violent content. However, performance is affected by the quality and variability of the training data. The study recommends further improvements through dataset expansion, integration of additional modalities, and comparative evaluation with other detection techniques.

Keywords: 3D- Convolutional Neural Network, Gaussian Mixture Variational Autoencoder, Video Violence Detection, Violent Video, Non – violent Video, Object Detection, Action Recognition