Inspiration
There have been countless mass casualty incidents as a result of failed crowd control and overcrowding in densely populated areas and gatherings. From the Travis Scott concert stampede in Houston, to Seoul’s Halloween-goers jam packed until they couldn’t breathe.
What it does
Our project provides a working solution to a pervasive problem. Utilizing an artificial intelligence-based solution, we are INNOVATING WITH AI. We are taking the first bold step to SeeF1rst, leading to preventive crowd control by calculating population density from image feed using YOLO object detection in python. We trained our own model using custom parametrizations to output the image feed and its respective population density.
How we built it
In the backend, our model uses YOLO to detect bodies when fed an image. This was done by training the model to find a dataset of human bodies and faces, convert the dataset to YOLO, and run 15 epochs with 15,000 images, a total of 10.7 GBs of raw data. This allowed our model to match the validation data with around 70% accuracy after training it overnight.
Challenges we ran into
Along the way we had issues with the data format when it came to training our model. In order to work around this issue were forced to write a function that reformats the image labels.
Accomplishments that we're proud of
We were able to train a working model which is able to detect humans in an area with a 70% accuracy.
What we learned
We learned fundamentals of project building which are not learned through traditional courses.
What's next for SeeF1rst
We will continue to work on the program to make it more streamlined.
Log in or sign up for Devpost to join the conversation.