AI-Enabled Smart Cane for Blind Individuals

Abstract

This project identifies design, implementation, and evaluation of a Smart Object Detection Blind Stick- an assistive mobility aid that will be used to complement traditional white canes by providing innovative obstacle detection and optional navigation on a worldwide scale that is based on GPS. The system overcomes the inherent weakness of traditional white canes that are only sensitive to objects in their immediate physical proximity, hence the user is exposed to obstacles in the air, those in the middle of the road, and those a long distance away. The gadget includes a slim YOLOv5n object detector model transformed into Tensor Flow Lite, which is installed on Raspberry Pi 4 hardware to run without relying on the internet connection. Public datasets of Roboflow and COCO were used to train the model with additional images of university furniture collected on-campus. The system has sensor fusion which is the combination of visual object recognition using a Pi camera and accu rate distance detection using an HC-SR04 ultrasonic sensor (2-400 cm range). The audio feedback is presented to the user through wired earphones using an offline text-to-speech engine. A optional GPS navigation mode uses Neo-6M GPS and QMC5883L compass modules to provide turn-by-turn direction guidance and the mode is mutually exclusive since the computational task was considered too complex. Some of the capabilities and limitations were identified during system evaluation. The prototype has 4-8 frames and 50-55 percent detection confidence. The performance is higher in the outdoor (70-80% detection rate) than the indoor (55-60%) because of improved lighting. On 10,000 mAh power bank, battery life was 7-8 hours. Nonetheless, major drawbacks are poor performance at low-light (30-40% detection rate), thermal regulation above 35°C, and GPS resolution of +-5 meters. Most importantly, there was no actual user testing on the visually impaired. The project shows technical feasibility of embedded AI-powered assistive devices under the budget constraint condition (PKR 40,000) and clearly records the vast difference between the capabilities of the current prototype and specifications to deploy assistive technology reliably