Inspiration
When brainstorming, we interviewed our community to gather input on what they felt needed to be addressed by technology. One such topic was given to us by two fire marshals with experience in dealing with the potentially deadly dangers of overcrowding.
What it does
Our app uses the Mask R-CNN model with the Detectron2 library to perform a semantic segmentation, pixel-by-pixel labeling of an image, to generate larger labels of areas (people). From there, we count the live number of people in a given area and determine whether or not the current density should be addressed before the situation becomes dangerous.
How we built it
Fundamentally, this is an object detection task. We toyed with the idea of training a large-scale object detection model from scratch but ultimately realized that doing so would be computationally intractable during a hackathon. As such, we opted to leverage powerful, well-researched, and battle-tested publicly available models.
Challenges we ran into
At first, we attempted to apply YOLOV8, a common model used for general object detection based on a CNN architecture, capable of supplying near-real-time object detection. As a result, we had trouble distinguishing an individual from their background, a potentially fatal mistake when each person is needed. However, through further research, we were able to utilize a more powerful model and drastically improve our detection rate.
Accomplishments that we're proud of
We are extremely proud of our ability to stream a live video feed and analyze it using computer vision in real-time, allowing for efficient and effective detection of a group of people. We are also proud of our easy-to-use user interface that allows for a quick analysis of any given situation.
What we learned
We learned how to train and use appropriate models, such as Mask R-CNN, to effectively detect individuals.
What's next for Occupan-See
The next step for Occupan-See is to expand our systems into fields where they could save lives, such as the architecture design industry.
Log in or sign up for Devpost to join the conversation.