Recently, bullying and other incidents occur frequently in primary schools. How to improve the campus safety through artificial intelligence technology has been one of the hot research topics. This paper investigates the intelligent identification method of risk behavior in primary schools based on audio and video fusion method. Firstly, we construct the video and audio data sets of risk behaviors in primary schools. Then, an intelligent recognition model of risk behavior based on Yolov5 algorithm is proposed. Next, the OpenAI-Whisper model is used for translation of audio data and accurate matching and recognition according to dirty word database to realize the intelligent classification of speech bullying. Finally, experiment results are given to illustrate that the proposed method can be used to intelligently identify bullying behaviors of students with high performance in primary schools, which can help school administrators and teachers take more effective prevention and response measures to improve the safety level of the campus.