Radar is often described as the "Goldilocks" sensor for autonomous systems. Unlike cameras, it preserves privacy and works in total darkness or thick smoke; unlike LiDAR, it is inexpensive and robust to adverse weather. However, radar data is notoriously "messy." The primary culprit behind this mess is a phenomenon known as multipath propagation, which creates "ghost targets"—detections of objects that do not exist in physical space.
The Geometry of a Ghost
To understand why ghosts appear, we must look at how radar measures distance. A radar sensor emits an electromagnetic wave and measures the time it takes to return. In a "clean" environment, the wave hits an object and bounces directly back. The sensor calculates the distance based on this round-trip time.
In complex environments like tunnels, shipping terminals, or indoor rooms, the signal rarely takes a single path. As Ryoma Yataka and colleagues explain in RETR: Multi-View Radar Detection Transformer for Indoor Perception, signals frequently bounce off secondary surfaces—like a metal wall or a polished floor—before returning to the sensor. Because the signal traveled a longer, indirect path, the radar processor perceives the object as being much further away than it actually is, often placing a "ghost" detection behind a wall or beneath the floor.
These artifacts are not just random noise; they are coherent reflections that mimic real objects. In industrial settings filled with metallic structures, Felix Fent notes in the MAN TruckScenes dataset that these reflections can be so severe they obscure actual obstacles, leading to false positives that can trigger unnecessary emergency braking.
Distinguishing Real from Reflection
Modern research focuses on three primary strategies to "exorcise" these ghosts: temporal consistency, cross-modal distillation, and environmental context.
Temporal Consistency
Ghosts are often transient or move in ways that defy physical logic. While a real car moves along a predictable trajectory, a ghost reflected off a curved guardrail might appear to teleport or move at double the ego-vehicle's speed. SIRA (Scalable Inter-frame Relation and Association for Radar Perception) addresses this by looking across longer temporal horizons. By using motion-consistency modules, the system can identify "pseudo-tracklets" that don't match the physical constraints of real objects. Similarly, Peizhao Li demonstrates in Exploiting Temporal Relations on Radar Perception for Autonomous Driving that leveraging the fact that physical objects maintain consistent properties over time allows models to filter out multipath interference that only appears in specific frames.
Cross-Modal Distillation
If radar is "messy" but LiDAR is "clean," why not use LiDAR to teach the radar what a real object looks like? This is the intuition behind RadarDistill by Geonho Bang. During training, a LiDAR "teacher" network provides high-fidelity structural features. The radar "student" learns to ignore the points in its cloud that the LiDAR indicates are empty space, effectively learning to suppress ghost reflections. Yizhou Wang 0005 pioneered a similar "teacher-student" approach with RODNet, using camera-radar fusion to automatically label radar data and filter out phantom detections in reflective environments like tunnels.
Environmental Context
Sometimes, the best way to identify a reflection is to know where the "mirrors" are. In mmWave Radar-Based Non-Line-of-Sight Pedestrian Localization, Byeonggyu Park uses a camera to map the road layout. By knowing the position of the walls, the system can mathematically "unfold" the radar signal, realizing that a detection appearing behind a wall is actually a reflected signal from a pedestrian around the corner. On the flip side, simpler heuristics can also be effective: Emerson Sie suggests in Radarize that prioritizing the first reflection along a given bearing can eliminate many secondary multipath ghosts.
From Points to Images
A more radical approach treats radar not as a collection of points, but as a noisy image. Junfeng Guan and Haitham Hassanieh developed HawkEye, which uses Generative Adversarial Networks (GANs) to perform image-to-image translation. The GAN is trained to take sparse, artifact-heavy radar heatmaps and "reconstruct" them into high-resolution depth maps. By learning the underlying distribution of how electromagnetic waves scatter, these generative models can effectively "hallucinate away" the ghosts, producing a clean environmental representation even in dense fog or rain.
As these techniques mature, radar is moving from a simple ranging tool to a high-fidelity imaging sensor. By turning the "weakness" of multipath reflections into a solvable geometric and machine learning problem, researchers are enabling autonomous systems to see clearly in environments where they were previously blind.