Semantic Segmentation on 3D Occupancy Grids for Automotive Radar

Abstract

Radar sensors have great advantages over other sensors in estimating the motion states of moving objects, because they detect velocity components within one measurement cycle. Moreover, numerous successes have already been achieved regarding the classification of such objects. However, the advantage of instantaneous velocity measurement is lost when detecting static objects, so that their classification is much more demanding. In this paper, we use semantic segmentation networks to distinguish between frequently occurring infrastructure objects. The resulting semantic grids provide a location-based classification of the vehicle environment. Since even modern radars have a significantly poorer angular resolution than lidars, the relatively thin radar point cloud is accumulated in advance and transformed into 2D or 3D grids that act as network inputs. Occupancy grids are particularly advantageous here, since they calculate not only the obstacles but also the free spaces. With suitable parameter selection, which is very challenging due to the complexity of radar measurement, the resulting grids allow for good association with camera images. Finally, in order to evaluate possible advantages of 3D grids as network input with respect to the segmentation result, we created and evaluated a simulation dataset and two different real-world datasets in car parks and on motorways. As a result, Jaccard coefficients between 81% and 88% were achieved, depending on the dataset. It was also found that 3D input images lead to improvements in the car park dataset.

Multi-Class Road User Detection With 3+1D Radar in the View-of-Delft Dataset

“…Only few works have used 3+1D radars for object detection. In [29] the authors applied such a sensor to build a static 3D occupancy map of highway and parking lot scenes after filtering out dynamic targets. Afterward, the map is semantically segmented by image segmentation networks into the street, curbstone, fence, barrier, or parked car classes.…”

“…Its most trivial use is to distinguish static and dynamic objects after ego-motion compensation. E.g., while some research only keeps static radar targets [29]- [31], others use the Doppler information to keep only moving reflections to detect dynamic objects [3], [23], [32]. After first clustering the radar point cloud to generate object proposals, basic statistical properties (mean, deviation, etc.)…”

A Bayesian Framework for Integrated Deep Metric Learning and Tracking of Vulnerable Road Users Using Automotive Radars

“…Additionally, only macro-movements are considered in the simulator setup. These properties make it a common choice for use cases like occupancy grid generation in automotive radar [58].…”

Raw High-Definition Radar for Multi-Task Learning

“…The entire RAD tensor is considered for multi-view segmentation in [30]. Radar point cloud segmentation has also been explored to estimate bird-eye-view occupancy grids, either for LD [22], [39] or HD [34], [33], [37] radars.…”