Gesture-Based Drone Control Using Wearable Data and 1D CNN

Unmanned Aerial Vehicles (UAV) provide a means to access remote locations and perform tasks in dangerous zones, reducing operation costs and safety issues for workers. These aircraft are becoming increasingly prevalent for usages such as package delivery, surveillance, agriculture and entertainment, among others. However, the traditional remote controller is not intuitive and requires training to avoid safety issues, such as injuries and property damage, especially when handled by novice users. In this work, we propose a control system that is capable of recognizing a set of nine dynamic hand gestures, which are used as a natural alternative to execute UAV flight commands. To recognize the gestures we defined a 1D Convolutional Neural Network (CNN) architecture to classify time series of motion data, acquired in real-time using a custom wearable device with an MPU 6050 sensor. This wrist-worn wearable was used to build a new dataset consisting of 5500 gestures, which allowed to train and evaluate the CNN model. The proposed model performed with an accuracy of 98.59% and average precision of 98.32%. To validate the control system, we built a web application that simulates the flight and control of a UAV through the proposed gesture-based commands.