Traditional automotive radar is an incredibly robust sensor: it tracks targets through dense fog, operates reliably in harsh weather, and directly measures instantaneous velocity via the Doppler effect. However, conventional millimeter-wave (mmWave) radar has a fundamental geometric limitation: it is virtually blind to height.
Because standard radar sensors utilize antenna arrays arranged primarily along a single horizontal axis, they compress the three-dimensional world into a two-dimensional bird's-eye view (BEV) plane consisting only of range and azimuth (angle). This hardware constraint introduces elevation ambiguity. To a conventional radar sensor, an overhead highway sign, a low-hanging bridge, and a stalled vehicle directly on the road surface generate returns that look identical in the 2D projection. As noted by Scheiner et al., this lack of vertical resolution makes it difficult for autonomous systems to differentiate between harmless overhead infrastructure and critical ground-level hazards.
Why Elevation Ambiguity Breaks Sensor Fusion
To build a reliable 3D spatial understanding of an environment, modern perception systems fuse radar data with other modalities like cameras or LiDAR. However, elevation ambiguity severely disrupts the geometric alignment required for sensor fusion.
When aligning a radar point cloud with a camera image, the system must map 2D radar returns into 3D space. Without native height data, assigning an arbitrary or fixed height to a radar point introduces severe spatial noise. As Wang et al. observe, this vertical uncertainty degrades 3D object detection accuracy, particularly for tall objects like trucks or buses.
Furthermore, this ambiguity complicates the geometric calibration between sensors. Luu et al. and Jiang and Saripalli demonstrate that estimating 6-DoF extrinsic parameters becomes mathematically ill-posed when the radar data lacks a clear vertical dimension, making it difficult to recover vertical translation and tilt parameters. In robotic localization and odometry, this limitation manifests as severe vertical drift over time, as shown by Kim in the development of Co-RaL.
Conventional Radar Return:
[ Range, Azimuth ] ---> Is it a bridge, a sign, or a car? (Elevation Ambiguity)
Visual Grounding Solution:
[ Radar Return ] + [ Vertical Search Window ] ──> Projected onto Image ──> Visual Context Resolves True Height
Strategy 1: Vertical Pillars and Visual Grounding
How can a network determine the height of an object if the radar hardware does not provide it? The core intuition is to use the visual context of a paired camera image to resolve the ambiguity.
Instead of treating a radar return as a single point at a fixed height, frameworks like CenterFusion (Nabati et al.) expand each sparse radar point into a vertical "pillar" or search window. When this pillar is projected onto the 2D image plane, it forms a vertical line segment that spans the plausible height range of objects in the scene.
A deep neural network can then scan along this vertical window. By leveraging the rich semantic and structural information from the camera, the network identifies the actual object surface (such as a vehicle) that matches the radar's range and velocity. As demonstrated by Singh et al., treating sparse radar data as vertical search windows allows the network to successfully map 2D sensor returns to their true 3D surfaces, effectively offloading vertical localization to the camera branch (Lei et al.).
Strategy 2: Cross-Modal Attention and Feature Repair
Rather than relying on explicit geometric expansions, more recent architectures use learnable attention mechanisms to bridge the dimensional gap between modalities.
In CramNet, Hwang et al. introduce a ray-constrained cross-attention mechanism. Because cameras lack depth and radar lacks elevation, the network uses the radar's precise range measurements to "pull" camera features along specific optical rays to their correct 3D locations. This constrains the search space and prevents the height ambiguity from muddying the fusion process.
Alternatively, when LiDAR is available, the fusion pipeline can operate bi-directionally. Wang et al. propose using dense LiDAR features to "repair" the radar's missing height information before performing final feature fusion. By supervising the radar features with LiDAR's precise 3D geometry, the network learns to predict the missing vertical structure, preventing the propagation of planar noise into the final detection layers.
The Horizon: Moving to 4D Imaging Radar
While algorithmic techniques like learnable densification (Huang et al., Lei et al.) and 3D offset prediction (Kumar et al.) offer powerful software-level corrections, they are ultimately bounded by the information limits of 2D radar hardware (Stäcker et al., Wang et al.). To resolve elevation ambiguity fundamentally at the hardware level, the field is transitioning toward 4D high-resolution imaging radar.
As noted by Yang et al. and Kim et al., 4D radar introduces vertical antenna arrays capable of measuring range, azimuth, Doppler velocity, and explicit elevation. Integrating these denser, height-aware point clouds remains an active frontier, promising to eliminate planar blindness entirely and unlock true 3D spatial mapping across all environmental conditions.
Key Foundational Work
RRPN: Radar Region Proposal Network for Object Detection in Autonomous Vehicles (2019) — contributed 1 concept elements to this direction.
RadarNet: Exploiting Radar for Robust Perception of Dynamic Objects (2020) — contributed 1 concept elements to this direction.