Automated multi-modal recognition of school violence

School bullying is a prevalent social problem worldwide, and the prevention of school bullying is an important and ongoing topic. Among all types of school bullying incidents, physical violence is considered the most damaging to teenagers. However, traditional anti-bullying methods are human-driven, which can be inconvenient for victims to use. This thesis therefore proposes automated multi-modal school violence recognition methods.

This thesis proposes sensor-based violence recognition solutions that can be implemented in wearable devices such as smartphones from victims’ perspective. Movement sensors integrated with an accelerometer and a gyroscope are used to collect the user’s movement data, and motion features are extracted to describe the differences between school violence activities and daily activities. With multi-modal information, the final recognition accuracy reached 93.6%.

From the perspective of teachers, this thesis proposes a video-based violence recognition solution. Each surveillance camera captures images of a specific area. The video-based solution achieved an average recognition accuracy of 97.6%.