Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Cone Detection with Convolutional Neural Networks for an Autonomous Formula Student Race Car
Submission Author:
Laíza Milena Scheid Parizotto , PR , Brazil
Co-Authors:
Laíza Milena Scheid Parizotto, Eric Aislan Antonelo
Presenter: Laíza Milena Scheid Parizotto
doi://10.26678/ABCM.COBEM2021.COB2021-0853
Abstract
Specific cones are used to delimit the tracks for the race cars in the Formula Student Driverless competition. To effectively race autonomously, the car must accurately detect them. In this context, this work investigates state-of-the-art Convolutional Neural Networks, specifically the so called You Only Look Once (YOLO) net, for robust and fast detection of cones from images of a camera mounted on the race car. To train YOLO, the Formula Student Objects in Context (FSOCO) dataset with four different classes of cones in the track is employed. The mean Average Precision (mAP) and network inference time are used to evaluate: (1) the influence of the image resolution; (2) the impact under different image perturbations such as brightness, exposure, blur, and noise; (3) the benefit of extra data augmentation for improving robustness to the aforementioned perturbations and out-of-distribution disturbances. We have found that: YOLO is a strong candidate for real-time cone detection for race cars; mAP increases with the image resolution, with just a slight increase in the network inference time; extra data augmentation for network training is beneficial for recovering the lost mAP when perturbations (of brightness, blur and noise) and especially out-of-distribution disturbances are applied to the images.
Keywords
Formula Student Driverless, Convolutional neural networks, Cone Detection, YOLO, Autonomous race car
