The CDC said in 2017 that the US saw 5,977 pedestrians killed by motor vehicles and more than 137,000 injuries. We can prevent these deaths using AI image recognition technology that is available right now. As a proof of concept, in just 24 hours on a personal computer with limited resources, I was able to train this human recognition AI with decent accuracy. If adopted by automobile manufacturers this technology can save lives.
What it does
This Haar image classifier is able to recognize a human from a simple video stream. This could be implemented in cars to apply driver warning systems or automatic breaking in order to prevent pedestrian motor vehicle accidents.
How we built it
Using OpenCV I was able to train a Haar classifier using thousands of images in order to detect a humans location in a video stream. This was all done on a consumer desktop windows machine within the span of 24 hours. If you would like to attempt something similar for yourself, the University of Auckland has a great tutorial that can be found here.
Challenges we ran into
Haar cascade training can be very resource intensive given it has to process thousands of images. It took about 9 hours for the Haar cascade implemented in this project to finishing training even when dedicated 8gb or RAM. The accuracy and reliability of these image classifiers can be greatly improved given more training input. While not very realistic on personal machine, these Haar cascades could be easily trained on enterprise systems with better resources and millions of images instead of thousands.
Accomplishments that we're proud of
While the low light accuracy of the image classifier is not the best, I am proud of how well this Haar cascade performs given it was developed in only 24 hours with not much experience.
What we learned
I've learned a ton about image recognition, python, and the theory behind Haar cascades. I look forward to applying this knowledge towards future endeavors!
What's next for Human Detector - HackAI
I would like to attempt to improve this Haar cascades accuracy by increasing the input images quantity and quality. Manually cropping these photos is very time consuming but a worthy use of time in my opinion to contribute to the future of AI, no matter how small that contribution is. Thanks for a fun HackAI!