Real time object localization based on computer vision: Cone detection for perception module of a racing car for Formula student driverless

Abstract

Using a Formula student car to race autonomously and find the right path has been around since the introduction of this class in the Formula student competition in 2017. This project work attempts this problem using a pure computer vision approach by utilizing a stereo vision camera only and running a convolutional neural network to localize the object in a field of view. To make sure the approach works at a stable high frequency, the detection of the cone-objects which are marking the path is run on a GPU. As the output of the system, the two dimensional coordinates in a local frame are calculated from each detection made in each frame, using merged approaches of triangulation and distance by size in the frame. The result is a solution with a sufficiently precise detection up to a distance around 12m, running at a frequency of 25Hz with more than 12 detected objects per frame, the system proved to be well enough designed to provide data for path planning module.