Research on Neural Network Technology in Public Safety Anomaly and Hazard Detection

Abstract

Abnormal hazards in public places have the characteristics of low frequency, diversity and strong timeliness. Single video model is easily affected by occlusion, illumination and network interruption. In this article, an abnormal danger detection method is constructed, which combines the temporal and spatial characteristics of video, environmental sensing information and communication status, and introduces the neural network synchronization control mechanism under DoS attack. The model uses convolutional network to extract local appearance and motion information, describes the evolution of events with gated time series units, and then completes multi-source feature fusion through attention mechanism. When DoS attacks cause intermittent loss of node data, state observation and synchronization compensation are introduced to maintain the time consistency between the edge detection results and the central model. The experimental results show that the F1 value of the proposed method reaches 0.922, which remains at 0.861 under the packet loss rate of 40%. Synchronous compensation reduces the average state error by about 41.7%. The research shows that the combination of dangerous content identification and network operation state can improve the stability of public safety anomaly detection under the condition of disturbed communication.