Aiming at the inefficiency and delayed response of traditional security management that relies mainly on manual monitoring and post-processing, a multimodal intelligent security management system is designed. The main components of the system include a visual recognition algorithm running on the Huawei Atlas 200I DK A2 development kit, a voice alarm device based on a single-chip microcomputer, and supporting software.Intelligent behavior recognition is achieved through visual processing algorithms and audio keyword detection.When dangerous situations occur, information can be automatically fed back to managers in time via the backend software, effectively ensuring on-site personal safety. For the visual algorithm, the YOLOv5 (You Only Look Once version 5) network structure is optimized by incorporating a CA( Coordinate Attention) mechanism to enhance detection capability for small targets and complex scenes, modify the loss function, and add support for the EIoU( Efficient IoU) loss function, enabling the model to adapt to scene changes and thereby achieve efficient recognition of fights and falls. Experimental results show that the mean average precision (mAP@ 0. 5)of the proposed method is improved significantly under various scenarios, and the detection speed meets real-time requirements, providing an intelligent solution for safety management in public places.

object detection, you only look once version 5 ( YOLOv5 ), computer vision, deep learning, safety management