In security applications, drones act as roaming sentinels that can cover blind spots and large perimeters more effectively than fixed cameraselistair.com. The challenge here is to use drone video or images for surveillance of a restricted area (e.g., a warehouse, border, or campus) and automatically detect unauthorized persons or vehicles. Drones provide an aerial vantage that can follow movement and respond quickly to intrusions. An effective solution could help security personnel by automatically spotting intruders in real-time and potentially tracking them.
Person/Vehicle Detection – Develop a vision model to detect human figures (and optionally vehicles) in the drone’s field of view. The drone may be patrolling an area that is supposed to be empty or only have authorized personnel. Any detected human or car is a potential “intruder.” The core task is to output bounding boxes (or simply an alert) when a person is present in the frame.
Zone Restriction – If the scenario includes specific zones (e.g., a drone monitoring a fenced property), participants can define a “virtual boundary” in the image and check if any person crosses into that area. This could be as simple as detecting if a person is present at all (if no one should be there), or if present outside of a designated safe zone.
Multi-Frame Tracking – For an extra challenge, track the intruder across frames. Once a person is detected, follow their movement in subsequent frames (this can be done by re-identifying the person’s bounding box in each frame). This isn’t required but demonstrates a more complete surveillance solution (tracking an intruder’s path).
Notification Output – The system should produce an alert when an intruder is detected. This could be a simple printout “Intruder at coordinates X,Y in frame 150” or drawing a bounding box on the video frame. Participants might simulate a siren by just highlighting the intruder in red on the footage. The key is to clearly indicate when/where an unauthorized entry was found.
Drone Human Detection Dataset – A dataset of drone images for human detection, containing 1,551 images annotated with human bounding boxes toolbox.google.com. This dataset (often available via AIHUB or Roboflow under titles like “Human Detection Through Drone”) provides overhead and oblique drone views with people labeled, which is ideal for training or testing your intruder detection model. Using these images, one can train a model to recognize human shapes from a drone’s perspective.
(Other options: the VisDrone dataset also contains person annotations in various scenes, and could be used for additional training data paperswithcode.com. If thermal imagery is of interest (for night detection), datasets like the UAV Human IR datasets could be mentioned, but for a one-day challenge sticking to RGB imagery is simplest.)
Detection Accuracy – The primary metric is how reliably the system spots an intruder when one appears. Missed detections (false negatives) are critical errors in security – a high-performing solution should have a very low miss rate for clear intruders in frame. Some false positives (thinking something is a person when it isn’t) are more tolerable than misses, but excessive false alarms will reduce the system’s practicality. We will likely test your model on a set of images (some containing people, some empty) and expect correct identification.
Speed – In a real security scenario, the system should work in real-time or near-real-time. Solutions that can process video frames quickly will be scored better. If you’re doing heavy computation (like running a deep neural network), consider the runtime: a solution that achieves slightly lower accuracy but at 5 FPS vs one that’s very accurate but 0.5 FPS might be judged as less practical for this context.
Simplicity & Reliability – Given the challenge scope, a straightforward approach that works consistently is better than an overly complex one that is unreliable. For example, if you implement a simple motion detection combined with person detection to double-confirm an intruder, and it works 9 times out of 10, that’s great. The judges will value a solution that is clearly explained and robust in its assumptions (e.g., “our solution assumes the area is empty, so any human detected is an intruder” – which is a fair assumption for this challenge).
Intruder Detection Code – A script or application that takes video/images as input and outputs an indication of intruders. This could output an annotated video or a series of image frames with boxes around detected people. It should be easy to run (with a sample input included).
Example Output Images/Video – Provide a few sample frames showing the drone’s view with an intruder highlighted (e.g., a red bounding box around the person). If the drone video is long, maybe include a short edited clip (5-10 seconds) demonstrating the detection in action.
Write-up – Describe your approach: e.g., “We used YOLOv5 trained on VisDrone to detect people. When a person is detected, we log the event. We define an intruder as any person detected since none should be present normally.” Mention the confidence thresholds or any logic (like if multiple frames detection is required to confirm). Also note performance – e.g., “processing ~3 frames per second on CPU”. This helps gauge feasibility.
Test Results – If you had a small test set (for instance, 20 images with and without people), you can report how many intruders were correctly detected vs false alarms. Even a qualitative assessment like “in our demo video, the system caught 9/10 intrusions, with 2 false alerts when the drone altitude changed rapidly” is useful.
Future Directions – Since security is a broad field, a short section on how you’d handle more complex cases: multiple intruders, distinguishing authorized personnel (maybe via uniforms or badges) from intruders, nighttime surveillance (thermal cameras), automated drone response (following the intruder), etc. This isn’t scored quantitatively, but it shows your understanding of the problem’s scope beyond the hackathon.