Development of a Gesture Control System for Intuitive Control of a Quadcopter
Abstract
This paper presents the development and implementation of a quadcopter control method using hand gestures within the ROS2 framework. The approach enables intuitive operator-drone interaction in environments where conventional remote controllers are unavailable or impractical. An Intel RealSense D435i depth camera stream provides precise spatial localization of hand keypoints. High-level control functions — motor arming and disarming, takeoff, and landing — are executed through predefined static gestures. Gesture classification is performed by analyzing the spatial arrangement of hand keypoints in the image plane using a fully connected multilayer perceptron comprising two hidden layers with 168 and 546 neurons, respectively, achieving an optimal trade-off between inference speed and accuracy. Continuous quadcopter control is accomplished by computing roll, pitch, yaw angles, and throttle values derived from three-dimensional hand position analysis, utilizing the camera’s intrinsic calibration parameters and depth map data. To mitigate gimbal lock effects near 90° angles, a deviation range constraint is incorporated into the control pipeline. Control signal smoothing employs a combination of a moving average filter with window size 5 and exponential smoothing with coefficient a = 0.5, ensuring rapid system response while effectively suppressing outliers caused by image processing artifacts. All filtering parameters were determined through systematic experimental evaluation. The system achieves gesture classification accuracy exceeding 99 %, with per-frame processing time below 33 ms (~30 FPS), which eliminates control signal latency and ensures high responsiveness. Test flights conducted in the Gazebo simulator using a DJI Tello quadcopter model validated the proposed method and confirmed its practical potential for interactive unmanned aerial vehicle control applications.