Deformable Radar Polygon: A Lightweight and Predictable Occupancy Representation for Short-range Collision Avoidance

TL;DR

The paper addresses short-range occupancy detection for autonomous driving by introducing a deformable radar polygon that encodes free space around the ego-vehicle using radar point clouds and Doppler information.It presents two algorithms—a basic radar polygon formation and a probabilistic, tracking-based update—to yield accurate and smooth polygons with low memory and computational requirements.A deformable polygon prediction capability uses vertex Doppler velocities to forecast near-future polygon shapes, enabling proactive collision avoidance and route planning.Extensive experiments on a self-collected dataset and RadarScenes demonstrate superior IoU-gt and IoU-smooth performance compared with baselines, along with robust collision-detection capability and favorable runtime/memory profiles.

Abstract

Inferring the drivable area in a scene is crucial for ensuring a vehicle avoids obstacles and facilitates safe autonomous driving. In this paper, we concentrate on detecting the instantaneous free space surrounding the ego vehicle, targeting short-range automotive applications. We introduce a novel polygon-based occupancy representation, where the interior signifies free space, and the exterior represents undrivable areas for the ego-vehicle. The radar polygon consists of vertices selected from point cloud measurements provided by radars, with each vertex incorporating Doppler velocity information from automotive radars. This information indicates the movement of the vertex along the radial direction. This characteristic allows for the prediction of the shape of future radar polygons, leading to its designation as a ``deformable radar polygon". We propose two approaches to leverage noisy radar measurements for producing accurate and smooth radar polygons. The first approach is a basic radar polygon formation algorithm, which independently selects polygon vertices for each frame, using SNR-based evidence for vertex fitness verification. The second approach is the radar polygon update algorithm, which employs a probabilistic and tracking-based mechanism to update the radar polygon over time, further enhancing accuracy and smoothness. To accommodate the unique radar polygon format, we also designed a collision detection method for short-range applications. Through extensive experiments and analysis on both a self-collected dataset and the open-source RadarScenes dataset, we demonstrate that our radar polygon algorithms achieve significantly higher IoU-gt and IoU-smooth values compared to other occupancy detection baselines, highlighting their accuracy and smoothness.

Figures (11)

• Figure 1: A high-level processing diagram for obtaining the radar point clouds from raw ADC data gao2023static.
• Figure 2: Procedures of forming a basic radar polygon. After sampling the FOV with $\Delta \theta$, a vertex (point ③) is selected based on normalized evidence probability, or a virtual vertex (point ②) is created for each sampling sector.
• Figure 3: The radar polygon update algorithm follows this processing diagram: First, the radar polygon from the previous frame $t-1$ is transformed to the new coordinate system centered at the radar at $t$, using the vehicle pose. The vertices in the transformed polygon are then integrated into the latest radar measurements for radar polygon generation. The formation steps follow the description in Section \ref{['sec:pol_ism']}. Specifically, there are four types of vertices: selected old vertex (point ①), selected vertex trackable to old vertex (point ②), uncertain emerging vertex (point ③), and virtual vertex (point ④). The uncertain emerging vertex will not be used for polygon formation in the current frame but will be stored and associated with future frames multiple times before being activated.
• Figure 4: (a) Example of a deformable radar polygon (i.e., polygon prediction); (b) Example of checking collision detection for point $(a, b)$.
• Figure 5: Left: Testbed with four radars (marked as blue rectangles), with their FOVs visualized in light green. Right: BEV parking lot where the car backing-off experiment was performed along the yellow trajectory (from Google Maps).