AI-Enabled Smart Walking Stick for Visually Impaired

Vision loss affects millions of people worldwide and creates major challenges in everyday mobility (Figure 1)[1]. Even simple tasks like walking down a sidewalk, crossing a street, or navigating indoors can become difficult and sometimes dangerous. Traditional canes are effective in detecting obstacles on the ground, but they have limitations (Figure 2). They cannot detect overhead obstacles, water on the ground (Figure 3) or provide advanced navigational assistance. This motivated me to pursue the development of AI-enabled smart walking stick. [2]

The purpose of my project is to enhance independence, safety, and confidence for visually impaired people by combining obstacle and water level detection with artificial intelligence for object recognition (Figure 4). With advancements in AI, sensors, and micro-controllers, it is possible to create affordable and portable systems that can provide guidance in real time [3][4]. However, many existing smart canes lack advanced features like object recognition, are expensive and not widely accessible [5].

This project aims to address these issues by designing a unique, low-cost, user-friendly walking stick that improves environmental awareness. The idea is not to replace the traditional cane but to enhance it with intelligent features. By doing so, users can benefit from both semantic feedback and environmental detection capabilities [6].

Ultimately, this project is about improving quality of life for visually impaired people. This smart walking stick has the potential to assist users and reduce reliance on others. This project shows how engineering and AI can be applied meaningfully to create a smart and innovative solution.

Background Research - Three main questions:

What is AI recognition and how does it work? (Figure 5)[7]

How different sensors help detect obstacles, and what are their limitations? (Figure 6)[4][8]

Are there existing smart canes, and what limitations do they have? (Figure 7)[5]

Material:

The design process began by identifying key user needs: obstacle and water detection, object recognition and clear feedback. A prototype was built using an ultrasonic and water level sensor, camera module, Raspberry Pi 5, Arduino Nano, speaker, buzzer, and a portable battery.

Design Process:

The system was programmed in stages. The ultrasonic sensor detects nearby objects, the water level sensor detects water on the ground, and the AI model on the Raspberry Pi processes camera input to classify objects and output it through a mini speaker (Figure 8). The Arduino Nano provides immediate alerts through sound signals for close obstacles or water hazards (Figure 9).

Data Collection:

A structured data collection process was conducted across three different environments: indoor, pathway and park.

For obstacle detection, 5 trials each for 3 different objects, were conducted in all 3 environment settings. Obstacles were placed at different distances. Data was collected by recording whether the obstacle was detected or not within 1 meter.

For water level detection, 5 trials each for 3 different water levels on the ground, were conducted in all 3 environments. Data was collected by recording whether water was detected or not by the sensor.

For object recognition, 5 trials each for 3 different objects, were conducted in all 3 environments settings. Data was collected by recording whether the objects were correctly identified or not within 1.75 meters.

The height and angle of the sensor were kept constant during trials, and tests were repeated under similar environmental conditions in each location to reduce variability.

Prototype Working:

On Raspberry Pi 5 system, the walking stick provides feedback through a multi-step process. The ultrasonic sensor continuously scans for objects, and if something is detected within 1.75 meters, activates the AI system. The camera sends visual data to the Raspberry Pi 5, where the AI model classifies the object and outputs the result through a speaker (Figure 10).

The Arduino Nano system provides instant danger alerts. If an object is detected within 1 meter, buzzer sounds at 320 Hz. If a water puddle is detected on the ground, the buzzer sounds at 500 Hz (Figure 11).

Results:

Obstacle Detection:

In the indoors, system was tested on a person, a chair, and a staircase. It detected all the obstacles 5/5 times within 1 meter and buzz. It did not detect any obstacles beyond 1 meter.

On pathway, the system was tested on a person, a car and a dog. It was able to accurately detect all the obstacles 5/5 times within 1 meter and nothing beyond 1 meter.

In the park, system was tested on a person, a tree and a bench. It was able to detect all the obstacles 5/5 times within 1 meter, and nothing beyond 1 meter (Figure 12).

Water Level Detection:

In the indoors, system was tested on dry surface, wet surface and 1cm of water in a bowl. It detected water 0/5 times on dry and wet surface, and detected water 5/5 times in the bowl.

On pathway, system was tested on dry surface, wet surface and water puddle. It detected water 0/5 times on the dry and wet surface and detected water 5/5 times in the puddle.

In the park, system was tested on grass, moist sand and water puddle. Out of 5 trials, it detected water 2/5 times on grass, 3/5 times on moist sand and 5/5 times in a water puddle (Figure 13).

Object Recognition:

Indoors, system was tested on a person, a chair and a TV. It was able to detect the person 4/5 times, the chair 5/5 times and the TV 4/5 times.

On pathway, system was tested on a person, a car and a stop sign. It accurately detected person 4/5 times, car 4/5 times and stop sign 3/5 times.

In the park, the system was tested on a bicycle, a dog and a bench. It was able to accurately detect bicycle 4/5 times, the dog 3/5 times and the bench 2/5 times (Figure 14).

Expert consultation from the Sight Loss community, feedback received:

The walking stick should not be bulky

The walking stick might cause the user to get overloaded by the sensory information, blocking out the sound

The water level sensor might be affected in the rain and snow

The angle of the walking stick could also cause issues with the sensor accuracy as visually impaired people use a walking stick with a sweeping motion

Conclusions:

The results show that the AI-enabled smart walking stick can effectively improve obstacle awareness compared to traditional canes. Indoor and outdoor testing showed high accuracy in recognizing objects (Figure 15) and reliably detecting obstacles (Figure 16), confirming that the combination of sensors and AI can provide fast, meaningful feedback. This supports the conclusion that integrating real-time object and water level detection with object recognition can enhance user safety and independence (Figure 17).

However, the results also showed important limitations.

Outdoor testing revealed challenges with environmental factors such as water, cold temperatures, and battery life. For example, water level sensor sometimes made continuous alerts after exposure to moisture, and system shutdowns occurred after 10–15 minutes.

Outdoor testing also showed a slight struggle with the AI model to correctly recognize all the objects in the environment.

These findings showed that while the prototype is functional, it requires improvements for better reliability in real-world conditions.

Learning:

From this project, I learned that building an effective smart walking stick is not only about adding advanced technology, but also ensuring consistency, and usability in different environments (Figure 18). Small factors like sensor placement, weather conditions, and power management can significantly impact performance.

Overall, the project demonstrates that AI-enabled assistive devices have a strong potential to improve navigation for visually impaired individuals and make them more confident while doing so (Figure 19). With improvements like better waterproofing, optimized power systems, and expanded AI training, the system could become a practical solution for everyday use.

This project could be extended by improving reliability and usability in real-world conditions:

Waterproofing and cold-proofing the sensors would help prevent false alerts (Figure 20).

Better power source or efficient micro-controller (Figure 21).

Improved AI model for better and accurate recognition (Figure 22).

Adding GPS navigation and voice directions (Figure 23).

Using feedback from expert from the Sight Loss community for future improvements (Figure 24):

Walking stick should not be bulky

Overload of sensory information

Water level sensor might be affected by rain and snow

Angle of walking stick during sweeping motion might cause ultrasonic sensor and camera unreliability

I would like to acknowledge multiple people for their feedback and support for my project.

First, I would like to acknowledge Mrs. Davis, Mrs. Fourie, Mrs. Tanner, and Mrs. Turner who are some of the teachers at Louis Riel School that helped me set up and get a start on my project.

I would like to acknowledge all the feedback and support from our CYSF delegates: Beth, Shannon, Sri, Scott, and Alex for help in improving my project for CWSF.

Huge thanks to an anonymous individual from the Sight Loss community for giving me feedback on my walking stick from their own personal experience as a visually impaired person.

Finally, I would like to acknowledge Gaurav Bansal (my dad), a software engineer, who was my expert, and who helped me with some code integration and who supported me during my time working on the project.

Information:

[1] World Health Organization. (2023). Vision impairment and blindness.

https://www.who.int/news-room/fact-sheets/detail/blindness-and-visual-impairment

[2] Sant, S., Salunkhe, M., Shah, K., Kumawat, G., & Palinje, M. (2023). Smart cane: A machine learning-powered assistive device. International Journal for Research in Applied Science and Engineering Technology, 11(4), 3815–3822. https://www.ijraset.com/research-paper/smart-cane-a-machine-learning-powered-assistive-device

[3] IBM. (n.d.). What is image recognition?

https://www.ibm.com/topics/image-recognition

[4] NVIDIA. (2020). How does a self-driving car see?

https://blogs.nvidia.com/blog/how-does-a-self-driving-car-see/

[5] WeWALK. (n.d.). WeWALK smart cane.

https://wewalk.io

https://www.tdk.com/en/featured_stories/entry_084-WeWALK-Smart-Cane-2.html

[6] Strickland, M. (2026). Semantic analysis of feedback: Transforming customer insights into actionable strategies. Luth Research https://luthresearch.com/glossary/semantic-analysis-of-feedback-transforming-customer-insights-into-actionable-strategies/

[7] Shaip. (2025, July 1). What is AI image recognition and how does it work? https://www.shaip.com/blog/what-is-ai-image-recognition-and-how-does-it-work/

[8] SparkFun Electronics. (n.d.). Qwiic ultrasonic distance sensor (HC-SR04) hookup guide. Retrieved April 26, 2026, from https://learn.sparkfun.com/tutorials/qwiic-ultrasonic-distance-sensor-hc-sr04-hookup-guide/all

[9] Mangrulkar, J., Bagde, H., More, R., Dohe, R., & Sakharkar, V. (n.d.). Empowering the visually impaired: Intelligent assistive device for blind people. Journal of Technology, 15(2) https://technologyjournal.net/wp-content/uploads/2-JOT1477.pdf

[10] EdjeElectronics. (n.d.). Train and deploy YOLO models [GitHub repository]. GitHub. https://github.com/EdjeElectronics/Train-and-Deploy-YOLO-Models/tree/main

Images:

Phoenix medical system cane: https://www.phoenixmedicalsystems.com/assistive-technology/smartcane

WeWalk Smart Cane 2: https://wewalk.io/en/

Spancare Smart Cane Walking Stick Electronic Aid: https://www.amazon.ca/Spancare-Smart-Walking-Electronic-Travel/dp/B0BWTRRXBK

Why?: https://firebasestorage.googleapis.com/v0/b/project-leo-mvp.appspot.com/o/attachments%2F797778be-7881-4fb6-864b-4f6c5b13835f%2F01e375bcd91?alt=media&token=6046b80a-73f7-49ef-bb56-7b9f63708ca5

How?: https://firebasestorage.googleapis.com/v0/b/project-leo-mvp.appspot.com/o/attachments%2F797778be-7881-4fb6-864b-4f6c5b13835f%2F160c95bcd91?alt=media&token=65e5eb13-6abc-4d93-a915-fae6b7e5abf5

What?: https://firebasestorage.googleapis.com/v0/b/project-leo-mvp.appspot.com/o/attachments%2F797778be-7881-4fb6-864b-4f6c5b13835f%2F524ae77cd91?alt=media&token=2e924b06-2f65-4e2a-9f55-adb8dcc18368

So What?: https://firebasestorage.googleapis.com/v0/b/project-leo-mvp.appspot.com/o/attachments%2F797778be-7881-4fb6-864b-4f6c5b13835f%2F5904f77cd91?alt=media&token=908e8cca-a792-4816-bcd1-ee24abb47638

What Next?: https://firebasestorage.googleapis.com/v0/b/project-leo-mvp.appspot.com/o/attachments%2F797778be-7881-4fb6-864b-4f6c5b13835f%2F478ff77cd91?alt=media&token=d4ec6865-4834-4447-a531-55f16379a881

Thanks: https://firebasestorage.googleapis.com/v0/b/project-leo-mvp.appspot.com/o/attachments%2F797778be-7881-4fb6-864b-4f6c5b13835f%2Fe807587cd91?alt=media&token=26ecf4de-4c16-44b7-9972-28ac5253535a

References: https://firebasestorage.googleapis.com/v0/b/project-leo-mvp.appspot.com/o/attachments%2F797778be-7881-4fb6-864b-4f6c5b13835f%2F2d0d687cd91?alt=media&token=65e89e84-3d88-45e6-88cb-dc9b86d4e298

OpenAI. (2026). Image generated by Bansal, Yugam using ChatGPT (DALL·E) [AI-generated image]: https://chat.openai.com/