Virtual Eye: Object Detection APP for Visually Impaired Presidential AI Challenge Track II Video

Описание: My inspiration for Virtual Eye (Traffic_Detection_App), an unique, accessible, accurate, and AI powered obstacle detection app, comes from my everyday instances: seeing visually impaired grandmothers waving their canes at non existent crossroad marks, hearing my neighbors complaining about the ever rising cost of a roadside service dog (up to $50,000!), and feeling the rough, kiss-the-wall-and-come-back consequences of not looking at the road myself (Dogster, 2024). Indeed, in the United States, more than six million people suffer from the daily obstacles of visual impairment, with one million of the population legally blind. Yet, this is only one tip of the iceberg: world-wide, almost 2.2 billion people are severely visually impaired, and it is estimated that there will be a doubling of the visually challenged community in 2050 (CDC, 2024). This community is almost constantly in danger: according to the International Journal of environmental research and public health, “68% of people with visual impairment had been directly exposed to at least one serious life event, with equal rates among males and females” (Brunes et. al, 2021). To service this part of America, and also the rest of the world, I am honored to introduce Virtual Eye, an YOLO v8 powered muti-platform obstacle detection application catered to suburban and urban street spaces for visually impaired populations.
I first used Yolo V8 (You Only Look Once Version 8), a state of the art AI-object detection and classification model to train on Google Collab (Python) with open source data sets from roboflow (exhibit 1). After this model reached 85% accuracy, we then transferred the trained model to TF-lite (otherwise known as flutter), an open source framework that offers space for presenting my technology to various platforms (IOS & Android).
The application set-up runs in two coding sites: Visual Studio Code (for android devices) and X-code (for IOS devices). When Virtual Eye is first runned, it asks for back camera permission, and once the permission is approved the screen opens up to two options—real time flow and manual capture. For real time flow, Yolo V8 first recognizes potential obstacles in a frame by frame manner, and then uses the roboflow open source data training to compare, analyze, and make its best guess. It then reports the best guess, or the classification that has the highest percentage, as an audible message. The range of objects that could be classified include: crosswalk, person, traffic light, vehicle, animal, side walk, stairs, pothole, train, tree, pole, and obstacle. When nothing potentially in the way is detected, the real time flow does not report. Manual capture’s logic is a repetition of real time flow, but this option gives more control to the user, allowing him or her to choose a specific scene to let Virtual Eye analyze (for example, traffic light).
A notable struggle I face while developing Virtual Eye is adjusting my macbook environment to be fit for developer mode. The downloading of visual studio and X-code took two shut downs and a total of 48 hours, and with the addition of unzipping cocoa pods, getting packages and linking dependencies, it took a whole week before I could actually start piecing models together and setting up the user interface. And as I communicated to various friends and experts in the developer and the business field, they all agreed that if there is anything that macbook could do to promote an fostering environment for potential young developers and future entrepreneurs, it is to make developer sites like X-code, Visual code, and Google Play Console easier to access and set up. However, this experience taught me the power of not giving up. I fought my inner doubts and stayed patient and steady throughout the process, focusing on unpacking issues and reaching out for help. The satisfaction I get after each “green light” report is enormous, and encourages me to keep going with my building.
As of now, I have already successfully made a google play console account, and have released a testing version of Virtual Eye for Android devices. Multiple testers have already downloaded the application in their smart phones and were reporting satisfactory results: they tested sidewalks, crossroads with moving cars and people, and the audio reports are always on time and accurate. The future for Virtual Eye holds great expectations: in the short term, I will add a separate model, training Virtual Eye to be sensitive to cross walk colors and to store more frames. In addition to these functions, I will design a more user friendly interface by offering customizable voices and range of objects to be detected. And in the long term, Virtual Eye will no longer be bound by the confines of a smart device: it will be included in smart glasses, classifying and reporting scenes at high speed and even higher accuracy.