Real-Time Object Detection and Classification for Enhanced Independent Navigation

Visually impaired individuals face significant challenges in navigating complex environments independently. This study proposes a lightweight, portable navigation aid system based on Raspberry Pi, integrating a camera, ultrasonic sensors, and the MobileNet deep learning framework for real-time object detection and classification. The system runs on the Raspbian OS and utilizes OpenCV and TensorFlow Lite to process visual data. A custom dataset of 600 images (person, chair, table) was used to fine-tune the MobileNet model, achieving a classification accuracy of 91.3%. The HC-SR05 ultrasonic sensor attained 99.06% accuracy in distance estimation, with an average deviation of 2.9 cm. The system delivers dual-mode validation (vision + distance) and provides audio feedback based on object type and estimated steps, calculated from the measured distance. It maintains a real-time processing speed of 18 frames per second with feedback latency between 0.4-0.7 seconds. Experimental tests confirm reliable performance across indoor conditions, demonstrating the system's potential as an effective, low-cost assistive technology. This research contributes a novel combination of deep learning and sensor fusion on an embedded platform, offering enhanced autonomy and spatial awareness for visually impaired users.