Abstract: In light of the high-risk challenges brought by the complex environment with limited space in cable tunnel construction and the problems of response lag and recognition blind spots in traditional manual monitoring methods, this study aims to integrate environmental parameters and personnel behavior monitoring by constructing a multi-source perception network and propose an improved YOLOv5s algorithm to enhance the recognition ability of safety equipment in low-light scenarios. Establish a multi-source information fusion risk determination model based on D-S evidence theory, design a dynamic risk classification early warning mechanism to link on-site prevention and control equipment, and develop intelligent risk prevention and control technologies to enhance the intrinsic safety level. Verification shows that the system has achieved a high-risk behavior recognition rate of over 94%, with a response speed increase of 60% compared to traditional methods. It has formed a closed-loop intelligent prevention and control system, providing an innovative solution for tunnel construction safety management.