Software for detecting potentially dangerous items at mass event entries using computer vision

The work is devoted to the development of software for automated detection of potentially dangerous objects in the form of large bags/backpacks on video streams at the entrance of participants to mass events. A client-server solution architecture is proposed, which includes a Streamlit interface, an Ultralytics YOLOv11 detection module, and an incident registration and alerting subsystem. The accuracy of bag identification is verified using the formalized «coordinated detection» criterion: an event is recorded when the confidence threshold γ is exceeded on at least α frames in a sliding window of length τ. On four RTSP channels (processing every 3rd frame), latency of ≈0.5 s was achieved, and false positives were mitigated through time aggregation. The YOLOv11 neural network was trained on a dataset containing 36,000 images, resulting in stable performance (P ≈ 0.953 for «dangerous» object detection accuracy and R ≈ 0.896 for completeness/sensitivity).