Testbed for Connected Artificial Intelligence using Unmanned Aerial Vehicles and Convolutional Pose Machines

TL;DR

This paper evaluates the efficiency of edge processing versus onboard UAV processing for real-time human pose estimation using Convolutional Pose Machines, demonstrating that edge processing is more efficient for CNN-based applications on UAVs.

Contribution

It provides an empirical comparison of processing constraints on UAVs versus edge devices for CNN-based pose estimation, highlighting the benefits of edge computing.

Findings

Edge processing reduces battery consumption.

Edge processing offers faster pose recognition.

Edge processing improves overall efficiency for UAV applications.

Abstract

Unmanned Aerial Vehicles (UAVs) became very popular in a vast number of applications in recent years, especially drones with computer vision functions enabled by on-board cameras and embedded systems. Many of them apply object detection using data collected by the integrated camera. However, several applications of real-time object detection rely on Convolutional Neural Networks (CNNs) which are computationally expensive and processing CNNs on a UAV platform is challenging (due to its limited battery life and limited processing power). To understand the effects of these issues, in this paper we evaluate the constraints and benefits of processing the whole data in the UAV versus in an edge computing device. We apply Convolutional Pose Machines (CPMs) known as OpenPose for the task of articulated pose estimation. We used this information to detect human gestures that are used as input to send commands to control the UAV. The experimental results using a real UAV indicate that the edge processing is more efficient and faster (w.r.t battery consumption and the delay in recognizing the human pose and the command given to the drone) than UAV processing and then could be more suitable for CNNs based applications.