A Multiclass Object and Depth Estimation Deep Model for Humanoids Robots in a Soccer Simulated Dataset

• Gabriel Previato UNICAMP
• Esther Colombini UNICAMP

Resumo

With the evolution of humanoid robotics and its increasing use in diverse environments and tasks, robots have to interact with the environment and, therefore, understand it accurately to execute decision making. In this work, we present the process of collecting a new dataset for simulated soccer scenes over the V-REP simulator using the NAO robot as the preferred robot architecture. For this purpose, we simulated a RoboCup soccer challenge scenario and collected over 200k images that contain up to 4 classes of objects with their corresponding categories (goal, teammate, opponent, and ball), depth estimation and bounding boxes. An extended multiclass version of J-MOD2 was trained to validate the dataset. Results showed that even when working with a more significant number of classes than the original setup, we achieved better results than the original structure. We also tested the algorithm for partial occlusion and modifications in the scene, such as the opponents color and ball size, and we found that the system was able to generalize for these modifications.

Palavras-chave: Computer architecture, Training, Estimation, Microprocessors, Humanoid robots