Inspiration
We have produced a weapons detection algorithm as we have personal connections to survivors of gun violence. Our college, St. Mary's College of Maryland, is minutes away from Great Mills High School, which experienced a shooting in 2018. Many of our close friends and peers were traumatized by the attack, and their experiences inspired us to develop a way to reduce the violence.
What it does
Our model uses instance segmentation to identify the presence and type of a weapon in a given image. The main use case we are tailoring towards is in areas of congregation, such an schools, malls, concerts, etc. By detecting weapons in CCTV footage, the model will be able to increase the effectiveness of security in these highly populated locations with automated identification of threats. It should be noted that this is not a prediction model, meaning it does not make assumptions on who may become a threat in the future, as these models are often heavily biased and unreliable. Instead, it identifies immediate threats for a quick response.
How we built it
We collected and annotated over 4,500 images to feed into our model using Roboflow. These annotations fell into one of five categories: Handgun, Rifle, Knife, Toy Gun, and null. With this information we used Roboflow to create an instance segmentation model that could then recognize these objects in new photos.
Challenges we ran into
We struggled to find enough photos of toy guns to populate our model. To try to remedy this, we doubled our training set for toy guns so that the model would have more info to learn from.
Accomplishments that we're proud of
We are very proud that we were able to train two models in the 24 hour time period, and get one model to have high accuracy. Additionally, the instance segmentation on our model came out beautifully, and the model is able to outline and identify the shape of the guns in the images.
What's next for Roboflow Weapons Detection Machine Learning Model
An important next step in this model will be getting the model to work on video, as the ideal application is CCTV footage. This will be a large undertaking and thus something we could not complete in the timeframe, but would take the project to the next level. Additionally, we would like the program to be able to keep a log of identifications and save the video in which it was used. This could be used to hold responders accountable for the way they handle the situations and identify disproportionate force.
Log in or sign up for Devpost to join the conversation.