Radar sensors are indispensable for autonomous systems operating in adverse weather conditions like fog, rain, or snow, where optical sensors fail. However, radar signals suffer from a unique and frustrating phenomenon: they see "ghosts."
Unlike cameras or LiDAR, which primarily rely on direct line-of-sight reflections, radar waves frequently bounce off multiple surfaces before returning to the receiver. Because radar hardware calculates the position of an object based on the total Time-of-Flight (ToF) of the radio wave, these indirect paths introduce severe spatial distortions. Understanding why these ghost targets form, why they evade traditional filters, and how modern machine learning either eliminates or exploits them is a central focus of robust perception research.
The Geometry of a Ghost
To understand how a ghost target forms, consider a standard automotive radar tracking a vehicle inside a concrete tunnel or an industrial shipping terminal. The radar emits a millimeter-wave (mmWave) pulse. Instead of traveling directly to the target vehicle and back, the pulse strikes the tunnel wall, bounces to the vehicle, reflects back to the wall, and finally returns to the radar sensor.
[Radar] -------------> (Wall)
\
\
v
[Vehicle]
Because the radar sensor assumes light travels in a straight line from the sensor to the object and back, it divides the total round-trip time by the speed of light. The resulting calculation places a virtual target—a "ghost point"—far beyond the actual physical boundaries of the environment, often appearing outside the drivable area or behind solid walls.
As Chenming He et al. demonstrate in their study of tunnel environments, these multipath reflections create a chaotic cloud of false detections. This problem is not unique to tunnels; it degrades performance anywhere highly reflective surfaces exist. For instance, Fangzhan Shi et al. document severe ghosting from indoor furniture when tracking drones, while Felix Fent et al. show that the massive metallic surfaces of shipping containers in logistics terminals generate overwhelming multipath noise.
Why Ghosts Fool Traditional Filters
In early radar processing pipelines, engineers attempted to filter out these artifacts using simple geometric heuristics or thresholding. However, modern dense environments reveal that ghost targets are highly sophisticated deceivers for two main reasons:
- Spatial Overlap: As Shanliang Yao notes, ghost targets do not neatly separate themselves from valid data. They frequently overlap with real physical objects in the spatial domain, meaning that naive geometric pruning risks deleting actual obstacles.
- Temporal Consistency: One might assume that as a sensor moves, a ghost reflection would instantly scatter and disappear. However, Ryoma Yataka points out that persistent multipath reflections can exhibit strict temporal consistency over multiple frames. If a ghost target moves in tandem with the real vehicle causing it, it forms a "pseudo-tracklet" that easily tricks temporal association algorithms into treating it as a legitimate moving object.
Furthermore, these artifacts violate the foundational assumption of scene flow and odometry algorithms: that points can be consistently tracked across time as rigid bodies. In frameworks like the self-supervised scene flow estimator RaFlow, multipath specularities act as non-rigid outliers that break standard motion constraints.
Learning to Denoise via Cross-Modality Distillation
If geometry and tracking consistency cannot reliably separate real targets from ghosts, how can a perception system clean up its radar data? The answer lies in cross-modality knowledge distillation.
LiDAR and cameras do not suffer from multipath ghosting in the same way radar does; light waves scatter differently and have much tighter beam widths. During the training phase of a neural network, researchers can use a well-behaved sensor (like LiDAR) as a "teacher" to train a radar-only "student" network.
In the RadarDistill framework, the network learns to map sparse, ghost-ridden radar point clouds to clean, dense LiDAR-like feature spaces. Because the loss function penalizes the network for predicting features where the LiDAR sees empty space, the radar network inherently learns to suppress activations caused by multipath echoes.
Similarly, Ruoyu Xu et al. leverage this cross-modality paradigm to help 4D radar models distinguish clean structural features from ghost points. By forcing the radar network's latent layers to match the spatial distribution of a LiDAR teacher, the student network implicitly learns the physics of multipath interference, suppressing the "virtual" returns while preserving the true physical structures.
Turning Ghosts into Features
While most automotive applications treat multipath interference as noise to be eliminated, an alternative paradigm seeks to embrace the physics of reflections to see around corners or through obstacles.
In Through-Wall Radar Imaging (TWRI), signals bouncing off multiple interior walls create severe ghosting that clutters the map. However, Honggao Deng et al. propose a clever inversion of this problem: instead of filtering out the double-bounce reflections, they incorporate them into a "double reflection imaging dictionary."
Because the relationship between a real wall and its ghost image is governed by deterministic optical geometry, the ghost target contains unique structural information about the layout of the building that a single direct reflection cannot provide. By treating ghosts as extra data points rather than noise, the authors demonstrate that multipath interference can actually enhance the reconstruction of hidden environments.
Whether suppressing ghosts through deep cross-modality distillation or exploiting them via geometric inversion, modern radar research has shifted from ignoring multipath artifacts to deeply modeling their behavior. This transition is what allows modern autonomous systems to maintain precise ego-motion and object detection, even when operating in the most chaotic, echo-prone environments.
Key Foundational Work
Building feature extraction method based on double reflection imaging dictionary (2019) — contributed 1 concept elements to this direction.
mID: Tracking and Identifying People with Millimeter Wave Radar (2019) — contributed 1 concept elements to this direction.
Radar-only ego-motion estimation in difficult settings via graph matching (2019) — contributed 1 concept elements to this direction.