In this study, a real-time pedestrian assistance system based on on-device processing is proposed to support the safety of pedestrians, including visually impaired individuals. The proposed system employs a lightweight YOLO object detection model to identify various obstacles such as crosswalks, traffic lights, electric scooters, bollards, and motorcycles, classifying them by risk level and delivering voice guidance to the user. In addition, a GPS-based distance measurement function provides pre-warning when approaching a crosswalk within a 30-m radius, thereby assisting spatial awareness. The system is designed to operate solely on an Android-based smartphone so that it can be used by anyone without additional equipment or training. Various experiments conducted in different pedestrian environments (daytime/nighttime, under diverse weather conditions) confirmed high object detection accuracy and revealed high user satisfaction in terms of intuitiveness and responsiveness. This study demonstrates the potential to enhance pedestrian safety and enable real-time mobility support in complex urban environments.