Simulation and Evaluation of Deep Learning Autoencoders for Image Compression in Multi-UAV Network Systems

Abstract

Mobile multi-robot systems are versatile alternatives for improving single-robot capacities in many applications, such as logistics, environmental monitoring, search and rescue, photogrammetry, etc. In this sense, this kind of system must have a reliable communication network between the vehicles, ensuring that information exchanged within the nodes has little losses. This work simulates and evaluates the use of autoencoders for image compression in a multi-UAV simulation with ROS and Gazebo for a generic surveillance application. The autoencoder model was developed with the Keras library, presenting good training and validation results, with training and validation accuracy of 70%, and a Peak Signal Noise Ratio (PSNR) of 40dB. The use of the CPU for the simulated UAVs for processing and sending compressed images through the network is 25% faster. The results showed that this compression methodology is a good choice for improving the system’s performance without losing too much information.