Investigating cooperative two-drone vision-based surveillance

Sammanfattning

This thesis investigates whether two low-cost camera-equipped drones can cooperatively perform vision-based semi-autonomous surveillance of a user-specified target. Scenarios such as search and rescue missions or crime scene documentation require both a close-up of the object of interest as well as an overview of the scene. With this in mind, a two-drone system was designed and implemented where a close-up drone searches for and tracks the object of interest while an overview drone follows the close-up drone to capture the whole scene. The system was first developed in the Gazebo simulation environment using the Robot Operating System 2. This simulation was then ported over to real hardware using Crazyflie 2.1 Brushless drones. The object detection for tracking was performed using YOLO26 Nano from Ultralytics. Bounding boxes from the object detection were translated into flight commands by three independent PID controllers, controlling the drones' yaw, altitude, and forward velocity. During real-life testing, the base YOLO26 Nano model met the 100-millisecond requirement (46 ms avg), but the custom-trained model did not (148 ms avg). The testing also showed an average correct-detection rate of 68.1%. The two-drone system met its core objective: the close-up drone semi-autonomously located and tracked a target while the overview drone maintained line of sight 98.98% of the time, demonstrating that cooperative vision-based surveillance is achievable on low-cost nano quadcopters using off-the-shelf components. Simulation-first proved a viable development path, and frame rate, detection performance, and battery life remain the main limitations.