Urban traffic congestion has become a pressing challenge in major cities, resulting in significant losses in time, cost, and residents’ quality of life. Computer vision techniques—particularly semantic segmentation—have demonstrated the capability to monitor and analyze traffic conditions based on real-time data using existing camera infrastructures, offering a practical approach to intelligent routing. This study proposes a traffic-density estimation method based on the DeepLabV3+ semantic segmentation model with two backbones, ResNet50 and MobileNetV2, trained on a manually annotated dataset collected from the public camera network in District 7, Ho Chi Minh City. Leveraging density information from the segmentation module, a closed-loop pipeline integrates an improved A* algorithm that incorporates both geographical and density-aware heuristics, enabling optimal route search based on current traffic flow. The system operates with periodic update cycles and low latency to ensure rapid response for navigation applications. Experimental results demonstrate that the segmentation models achieved high performance during validation, and the routing algorithm significantly reduces travel time across test scenarios, validating the effectiveness of combining computer vision with intelligent routing for urban traffic management.