Improving Object Classification Accuracy for Detecting Obstacles on Indoor Evacuation Paths Using YOLOv8

Purpose: This study proposes a foundational framework for estimating the effective evacuation path width in complex building corridors through real-time obstacle detection. The feasibility of a low-cost high-efficiency system was assessed by leveraging an existing closed-circuit television infrastructure to eliminate the need for additional hardware deployment. Method: A pretrained YOLOv8n model was used to detect objects on static frames from corridor surveillance footage. To address the challenges specific to indoor environments, a custom dataset was created via data augmentation and class relabeling using Roboflow. Performance was enhanced after training by applying class unification and parallel detection-based post-processing strategies. Result: After model customization and data augmentation, the system achieved improved detection metrics: precision of 0.9925, recall of 1.000, and mean average [email protected] of 0.9950. The results revealed the robustness of the model against occlusion and illumination variations and validate the technical feasibility of object detection as a core component of automated evacuation path recognition.