Abstract
Radar sensors employed for environment perception, e.g. in autonomousvehicles, output a lot of unwanted clutter. These points, for which nocorresponding real objects exist, are a major source of errors in followingprocessing steps like object detection or tracking. We therefore present twonovel neural network setups for identifying clutter. The input data, networkarchitectures and training configuration are adjusted specifically for thistask. Special attention is paid to the downsampling of point clouds composed ofmultiple sensor scans. In an extensive evaluation, the new setups displaysubstantially better performance than existing approaches. Because there is nosuitable public data set in which clutter is annotated, we design a method toautomatically generate the respective labels. By applying it to existing datawith object annotations and releasing its code, we effectively create the firstfreely available radar clutter data set representing real-world drivingscenarios. Code and instructions are accessible atwww.github.com/kopp-j/clutter-ds.