Abstract
Visually impaired individuals frequently encounter significant navigation challenges due to the limitations of conventional assistive tools, such as white canes or basic electronic aids, which typically provide reactive rather than predictive obstacle detection. These reactive systems issue alerts only upon the immediate proximity of obstacles, inadequately addressing dynamic environments where quick and proactive decision-making is critical. Consequently, users have insufficient time to avoid collisions or navigate safely, highlighting an urgent need for more advanced predictive assistive technologies.
This thesis addresses these challenges by developing an innovative predictive smart cane system integrating multiple advanced sensor technologies, including ultrasonic sensors and an Arducam OV5647 wide-angle camera, coordinated by a Raspberry Pi 4 Model B computing platform. Real-time obstacle detection is achieved using the YOLO-Fastest deep learning model, selected for its lightweight computational demands and real-time efficiency. This model precisely detects and tracks moving objects, analyzing positional changes between successive video frames to forecast short-term obstacle trajectories. Ultrasonic sensors complement visual predictions by continuously monitoring near-field ground-level anomalies, including sudden drops, stairs, potholes, and immediate obstacles. Validation and performance evaluation involved rigorous experimentation in both controlled environments and real-world scenarios, leveraging benchmark pedestrian trajectory datasets such as ETH-UCY and JAAD to assess predictive accuracy, responsiveness, and reliability.
The comprehensive evaluation demonstrated that the predictive smart cane significantly improves navigation safety and autonomy. The integrated multimodal sensor approach proved highly effective in enhancing predictive accuracy and environmental awareness, ensuring robust obstacle detection across diverse conditions. The system consistently provided proactive visual alerts, represented by directional arrows and trajectory markers within a clearly defined 'danger zone.' However, these visual indicators were used exclusively for system testing, validation, and demonstration purposes, helping observers and evaluators understand and visualize how the prediction mechanism functions. They were not intended for direct use by visually impaired users. Experimental results revealed high predictive accuracy, minimal latency (approximately 120 ms per frame), and superior responsiveness compared to existing reactive systems. This research validates the effectiveness of a predictive smart cane designed to enhance navigation capabilities for visually impaired individuals by integrating advanced sensor technologies and machine learning algorithms. Quantitative performance evaluations demonstrated superior predictive accuracy in static obstacle scenarios, achieving mean Euclidean errors between 2.05 and 5.17 pixels, indicative of high spatial precision in object localization. However, in dynamically challenging scenarios, such as fast-moving obstacles (mean Euclidean errors: 70.37–72.06 pixels) and sudden obstacle appearances (64.85–65.41 pixels), predictive performance notably diminished, reflecting inherent limitations of real-time object tracking and trajectory prediction using a single-camera embedded vision system. Despite these challenges, the developed system represents an affordable, accessible, and technologically robust solution that significantly enhances situational awareness, contributing substantially to improved independence, mobility, and overall quality of life for visually impaired users.