Emergencies such as natural disasters and military conflicts severely threaten the efficiency and safety of emergency logistics transportation. Traditional path planning research often neglects real-time road passability, resulting in planned paths that are often impractical in reality. This paper aims to construct an intelligent emergency logistics framework integrating “perception-decision”, achieving closed-loop optimization from environmental perception to dynamic decision-making by integrating deep learning-based road obstacle detection and operations research-based emergency rescue path planning. Firstly, a non-structured road obstacle detection system was developed based on the YOLOv5 algorithm, achieving a mean Average Precision (mAP) of 78% for four typical obstacle categories (Unobstructed, Pothole, Natural Disaster, Military Equipment) on a self-built dataset. Secondly, an emergency Vehicle Routing Problem (VRP) model aimed at maximizing the delivery volume of relief supplies was constructed and efficiently solved using the OR-Tools solver. Simulation and case analysis demonstrate that the model can achieve a 66% rescue efficiency under resource-constrained conditions. The core contribution of this paper lies in dynamically integrating real-time perceived road obstacle information into the path planning model, forming a collaborative optimization mechanism, thereby providing an effective solution for enhancing the intelligence level and mission success rate of emergency logistics systems in complex environments.