SafeSight

• 

  Goggles

Inspiration:

On the first day we focused on brainstorming as many problems we ourselves have faced. One of our team members had stated that, while working in construction, he noticed that it was very unsafe so we decided to further look into the data behind it. Turns out it is the leading industry in terms of death and injuries.

What it does:

AR goggles that provide real-time safety feedback on construction sites by highlighting dangerous objects. It also provides safety instructions while handling dangerous tools.

How we built it

As 48 hours is not enough to train an AI model we devised an architecture we will use. We chose to use MobileNetV2 to do the brunt of the work (used for detecting geometric shapes in an image) because it isn’t resource intensive which is a perfect use case for AR goggles which have limited processing power. Later this data is used for semantic segmentation (object recognition) which will be done in DeepLabV3. Later we plan on gathering data from construction sites (we have contacted a construction company) and training our model on these videos.

Challenges we ran into

There aren’t any trained AI models that are able to detect dangerous areas in construction sites. Also AR goggles have limited processing power so we had to find a way to make our AI model less intensive. As our market is relatively new, we were not able to find any direct competitors and had to conduct a wider competition research.

Accomplishments that we're proud of

During the 48 hours we have done extensive market research which proves that our product is necessary for the current market. We launched a successful survey that was aimed at our direct target audience, as well as interviewed an expert in AI technology and a work safety inspector who provided us with the most common issues in construction sites. Additionally, we contacted a nearby construction site who would let us gather the data necessary to train our AI. Lastly, we managed to build our first AI model that detects humans in videos

What we learned

Previously, we had never made an AI model used for object detection. We also further expanded our knowledge in doing market research as well as improved communication skills by interviewing specialists.

What's next for SafeSight

As stated before, we have found a connection that is eager to let us use their construction sites allowing us to gather data for our AI. Of course this wouldn't be the only company that we would need to cooperate with, but it is a start. After our AI is trained, we will start looking into producing our AR goggles. While the prototype itself with the projected shape is done, we still need to develop the needed technology for the AR tech to work, which shouldnt be the hardest part as similar technology already exists, but isnt used the way we will.

Built With

• deeplabv3
• mobilenetv2
• python
• torchvision