Offline Dynamic Grid Generation for Automotive Environment Perception Using Temporal Inference Methods

Abstract

The validation of online perception algorithms in automotive systems requires a large amount of ground-truth data. Since manual labeling is inefficient and error-prone, an automatic generation of accurate and reliable reference data is desirable. We present a post-processing approach based on a particle-based dynamic occupancy grid representation of the environment. In contrast to existing online dynamic grid algorithms, our estimation additionally utilizes future measurements by applying offline smoothing algorithms. Our proposed concept uses a two-filter procedure for smoothing the occupancy states of the grid cells. We further introduce two methods based on particle reweighting and two-filter smoothing to improve the velocity estimates. We show that our approach enhances the situational awareness and thus provides a more precise environment model. We demonstrate these benefits using lidar data from real-world experiments.