The research focused on developing a system to support the detection and classification of school violence based on video data and cameras. The research proposed two approaches: an indirect method of tracking object movement to identify the moving area, and a direct approach using classification from video.

Notably, the thesis developed the VSiSGU dataset, which contains detailed labels for school violence behaviours, providing a foundation for applying deep learning methods such as VGG16, LSTM, and GRU to improve accuracy in detecting and classifying school violence. This contributed effectively to preventing violence in educational environments.

The thesis presented several novel results in prestigious journals and scientific conferences, with highlights including a clear definition of school violence and the establishment of classification criteria based on images, alongside the development of the VSiSGU dataset specifically designed for school violence with reliable criteria.

The research also introduced improvements to the particle filter method, integrating motion direction and building an adjustment model to recover partially obscured objects during tracking. Furthermore, the study combined deep learning models (VGG16, LSTM, GRU) to enhance the accuracy of school violence classification with detailed hierarchical levels, and proposed a frame reduction method in videos to improve efficiency and practicality.

The development of an integrated approach combining object tracking algorithms with a regression neural network enabled accurate analysis of motion and classification of school violence behaviours.

These results were published in articles in reputable forums within the ISI and Scopus lists, affirming the research value of the thesis.

The thesis opened up several practical applications, including a school violence detection and alert system that could be integrated into educational institutions to reduce violence and improve management quality; the application of surveillance technology in other fields such as public security, detection of abnormal behaviour in traffic, and in crowded areas; data filtering and refinement to provide the user community with data, enhancing the feasibility of the system; and new research directions in building datasets and developing deep learning models to optimise behaviour detection in complex situations.

However, several issues remained to be addressed, such as improving accuracy under noisy data conditions, handling object occlusion, and developing models adaptable to diverse contexts.

Doctoral researcher Ngô Dương Hà’s thesis not only demonstrated her personal commitment to research but also made a significant contribution to raising awareness and providing solutions for school violence in modern educational environments.

ℙ𝕄ℕ