5) Crime prevention
We've heard many people, who lost their loved ones from a crossfire, a car accident, and much more. This would have never happened if the rescue came in time. This also would have never happened, if they were more aware. As people who enjoy computer vision, we decided to tackle this problem on unique way where only a few enterprises would utilize, and sell for high. If this program gets implemented for free, in a "text alert message" fashion, it will definitely let everyone be aware about their surroundings, and help the victims in the wrong place, wrong time to have another stay.
What it does
This object detection model detects with self made and transported labels, to detect fraud and criminal acts. This also predicts fires beforehand before it becomes serious.
How we built it
We used the State Of The Art YOLOV5 model to detect these images. Using advanced methods such as NMS, BFS, and soft NMS. Not only this, but we also hypertuned the model, even adding mosiac 1-4. WIth a threshold of 0.45. Using ONNYX and opencv to translate it into a tfjs model. There, we used inferencing with java's integrated camera library.
We also used React, JS for the frontend.
Challenges we ran into
Finding the dataset was definitely the challenge, and was definitely took the longest. Because we had to annotate them by ourselves, we later found an autoML to automate it for us(roboflow). However, there were more things that became a challenge.. My(Andy's), ROG gaming laptop 1650 GPU i5 9th gen CPU shut down(or exploded?) and did not ever turn up again. I(Andy), had to revert back to my old 2011 HP laptop which I almost forgotten how much I hated it. The Frames were slow, laggy, and memory usage was bottlenecking.
Accomplishments that we're proud of
We finally got it working! We have successfully worked until 3 am(MST) to finish this project, which was definitely a learning experience!
What we learned
Next time, don't waste time on finding good datasets to use, instead, follow the engineers code, "Make it barely work" ~Sun Tzu probably Then focus on good datasets to use and annotate.
What's next for SaS - Crime and Hazard Detection
Going to create a faster, webscaled, and lighter object detection model (using yolo5s.yaml) instead of the yolov5x.yaml. Definitely going to deploy it into a raspberry pi, and serve into more augmented reality environments!