Inspiration
After having witnessed the rising crime rates in cities like Brampton, Scarborough, and Vancouver, we brainstormed ways to reduce illegal activity. We needed a system for patrolling, a system revered by criminals, a system in the skies. We came up with Autopatrol, a non-intrusive, streamlined, and economic way to protect and dissuade.
What it does
During flight, Autopatrol will autonomously patrol areas to scan its surroundings and send this video feed to a computer. If it finds a dangerous object (e.g. gun or a knife), it reports to a security dashboard monitored by a safety contractor. The contractor can then decide if it’s important enough to warrant a 911 call, send some personnel to investigate, or if it is a false alarm.
How we built it
Autopatrol uses a DJI Mini 2 drone, with video data being sent to the phone as an intermediary. From there, an RTMP stream is sent from the phone and received on a computer by MediaMTX, which is then piped into OpenCV with a YOLO11 model designed to detect phones in place of weapons as an example (dataset credit: vladjkezor on RoboFlow). From here, it will send any detections to a security dashboard through a WebSocket. A security contractor will then decide if the situation requires a 911 call, and if yes, send a call to EMS (our devices) through Twilio. We used Google Antigravity to create a large amount of this pipeline, as well as the dashboard.
Challenges we ran into
Our main issue throughout this entire project was controlling the drone. Our original goal was to have fully autonomous control through either the DJI Mobile SDK or RosettaDrone. After many hours of work and several failed attempts, we were not able to successfully control the drone and decided to focus on other features.
Accomplishments that we're proud of
We managed to achieve around 1-3 seconds of latency with reliable AI image processing for object detection on a laptop with integrated graphics through clever use of multithreading. We were also proud of the fact that despite many setbacks, we were able to perservere.
What we learned
Overly proprietary and locked-down software/hardware, such as that of DJI, is very difficult to interface with, poorly documented, and severely outdated. We need to aim for more well-documented, popular, and up-to-date libraries in the future to accelerate development.
What's next for Autopatrol
We plan to upgrade our fleet with drones such as the DJI Matrice series which supports remote control SDKs. We can purchase our own servers for remote control and visual classification. We will reach out to security contractors to manage the dashboard and escalate if needed.
Log in or sign up for Devpost to join the conversation.