-
-
Our goal.
-
Our patrol map of the Bellot Strait.
-
Vision detection of the boat.
-
Upgraded vision detection.
What it does
Shadow Stalker is a program that searches for and detects "shadow vessels", sea craft that do not emit any Automatic Identification System (AIS) signals. These vessels are deliberately silenced in order to commit sabotage, conduct covert operations, and transport high-risk cargo.
How we built it
- Developed a custom computer vision model that outperforms YOLO in speed and versatility, processing video at up to 180 frames per minute
- Optimized paths and tower placements for coverage and efficiency by analyzing terrain constraints and mimicking lifeguarding search patterns
Challenges we ran into
There were so many other ideas we wanted to implement but had neither the time nor the resources to do. For instance:
- the flatness of the strait allowed a "cheese strategy" where a drone looking strictly horizontal at a 20m altitude would easily see the direction of the boat.
- additional computer vision strategies including a pixel-difference model to be used on a stationary quad drone to detect the boat from very long distances
- powering our fleet with a local LLM orchestrator to handle intercommunication and work around unexpected emergencies on the fly
Additional challenges:
- The high latency simulator environment limited the speed of our iteration, and we were also constantly set back by bugs, connection issues, etc. We also had to switch from cloud-hosted to local simulation environments as the cloud server broke down.
- The obligatory "it worked on my machine so it's fine" since we all had different operating systems
Accomplishments that we're proud of
- Being able to create a multi-layered vision engine that outperforms (in almost every way) the standard YOLO object detection model in this niche context
- Leonardo: my first all-nighter
- This is a project made by a team that genuinely found working on it interesting and fun (I dare say, more so than the events). No obligation to manufacture impact or meaning to the project idea, or to make it sound more "cool". This was the perfect challenge to work on, and we're proud that we could be part of it.
What we learned
- How to work with heuristics to optimize an open-ended problem
- Computer vision, search patterns, agents, mavproxy, 3D environments, defense
What's next for Shadow Stalker
- Mentioned above; many other ideas to try, and things to optimize for a practical environment that we couldn't get to in 24 hours
- Expanding our fleet to 4 drones to increase the range of distributive strategies and paths to explore
Log in or sign up for Devpost to join the conversation.