Embedded object detection and position estimation for RoboCup Small Size League

Abstract

In the RoboCup Small Size League (SSL), there is the challenge of giving more autonomy to the robots, so they can perform some tasks without receiving any external information. To achieve this autonomy, the robot has to detect and estimate the position of other objects on the field so it can score goals and move without colliding with other robots. Object detection models often use monocular images as the input, but calculating the relative position of an object given a monocular image is quite challenging as the image doesn’t have any information on the object’s distance. The main objective of this work is to propose a complete system to detect an object on the field and locate it using only a monocular image as the input. The first obstacle to producing a model to object detection in a specific context is to have a dataset labeling the desired classes. In RoboCup, some leagues already have more than one dataset to train and evaluate a model. Thus, this work presents an open-source dataset to be used as a benchmark for real-time object detection in SSL. Using this dataset, this work also presents a pipeline to train, deploy, and evaluate Convolutional Neural Networks (CNNs) models to detect objects in an embedded system. Combining this object detection model with the global position received from the SSL-Vision, this work proposes a Multilayer Perceptron (MLP) architecture to estimate the position of the objects giving just an image as the input. In the object detection dataset, the MobileNet v1 SSD achieves 44.88% AP for the three detected classes at 94 Frames Per Second (FPS) while running on a SSL robot. And the position estimator for a detected ball achieves a Root Mean Square Error (RMSE) of 34.88mm.