Automatic Detection System for Public Safety Events Combining Deep Neural Networks and Pattern Recognition

With the increasing complexity of public safety incidents, traditional incident detection methods are unable to meet the requirements of efficient and accurate detection, making the construction of automated detection system become particularly important. In this paper, we propose an automatic detection system for public safety events based on Vision Transformer (ViT), and optimize its performance through some innovative methods. Specifically, the model uses ViT to extract image features, and takes advantage of its long-distance dependency modeling to improve the spatial information understanding ability of events. In order to further optimize the timing modeling, a lightweight Temporal Convolution Module (TCM) is introduced, which can efficiently process the timing information of video frames and enhance the representation ability of timing features through an adaptive adjustment mechanism. In addition, combined with the multi-level Space-Time Self-Attention Mechanism, the dynamic weighting of spatio-temporal information is realized, which further improves the detection accuracy of the model in complex scenes. The experimental results show that the proposed method achieves higher accuracy and realtime performance than the existing technology on multiple public safety event datasets.