Drone Control through Deep Learning based on Pose Recognition

This research introduces an innovative approach to personal video capture using Unmanned Aerial Vehicles (UAVs) with pose recognition. The primary aim is to automate the video capture process for high-quality, self-captured videos. A deep learning algorithm was employed to develop a pose recognition model for real-time detection of an individual’s pose. This model was integrated into the UAV’s control system, enabling autonomous navigation. The trained model showed significant pose prediction accuracy (80.35%) within the training location with minor background changes. However, performance diminished under different perspectives (50.02%), locations (62.39%), and low-light conditions (29.44%). A prediction filter was implemented to enhance performance, enabling drone movement only after consecutive identical pose prediction. The findings demonstrate the effectiveness of pose recognition in capturing high-quality video footage autonomously. The integration of this technology in UAVs presents several advantages for self-capturing video, including unique perspectives and distinct angles without the need for an operator, leading to superior footage quality.
¹