A lightweight vision-only indoor navigation framework that constructs a room connectivity graph from a simple monocular video walkthrough and uses it for real-time guidance. Our approach harnesses vision-language models and classical computer vision techniques: we use CLIP image embeddings to recognize scenes, optical flow to estimate motion (detecting turns and transitions), and blur detection to filter out unclear frames. Each distinct room in the environment is automatically identified and represented by three key reference images (capturing the start, middle, and end views of the room), which serve as nodes in a topological graph. The graph encodes how rooms connect and the direction of movement between them (e.g. forward transitions, left or right turns). During live navigation, frames from a user’s smartphone or an IP camera are continuously matched against these stored references via CLIP-based similarity to localize the user’s position. A breadth-first search (BFS) on the room graph then finds the shortest path to the desired destination, and a text-to-speech module provides step-by-step audio instructions (for example, “Turn right to enter the corridor”). We evaluate our system on custom indoor video datasets: a fine-tuned CLIP ViT-B/32 model achieved an accuracy of 0.938 with an F1-score of 0.940, markedly higher than the ResNet’s 0.852 accuracy and 0.847 F1-score. Similarly, the ViT attained superior precision (0.912 vs 0.762) and slightly higher recall(0.971 vs 0.958).