When a sensor measures the physical world, its perspective dictates its blind spots. This is fundamentally true for velocity estimation. If you stand on a street corner and watch a car drive directly across your field of view, its speed is obvious. But if the car drives directly toward you, judging its speed becomes much harder because its lateral position doesn't change—only its apparent size does.
In engineering and computer vision, this creates a dual set of sensing blind spots: tangential blindness in frequency-based sensors (like radar and ultrasound) and radial blindness in visual sensors (like cameras). A major research direction focuses on resolving these geometric ambiguities to reconstruct full, multi-dimensional motion vectors.
The Geometry of Directional Blindness
To understand why this ambiguity occurs, we must look at the underlying physics of how different sensors measure motion.
A Doppler radar or ultrasound sensor measures velocity by detecting the frequency shift of a reflected wave. This shift is entirely governed by the relative motion along the sensor's line of sight. Mathematically, the measured scalar Doppler velocity $v_r$ is the dot product of the true 3D velocity vector $\mathbf{v}$ and the unit directional vector pointing from the sensor to the object $\mathbf{\hat{r}}$:
$$v_r = \mathbf{v} \cdot \mathbf{\hat{r}}$$
If an object moves perfectly perpendicular to the sensor beam, the true velocity vector and the directional vector are orthogonal ($\mathbf{v} \perp \mathbf{\hat{r}}$). The dot product drops to zero, resulting in a zero Doppler shift.
As Mian Guo explains, this means an object moving purely tangentially appears completely stationary to a Doppler sensor. In autonomous driving scenarios, this geometric limitation can cause a vehicle crossing an intersection to be dangerously misclassified as part of the static background landscape.
Tangential Motion (Blind Spot)
-------> v
^
|
| r̂ (Line of Sight)
|
[Sensor] --> Measures 0 Doppler Shift
Conversely, traditional 2D cameras suffer from the exact opposite problem: radial blindness. As established early on by Ramesh C. Jain et al., standard motion detection algorithms excel at tracking lateral displacement but struggle with axial movement (motion directly toward or away from the lens).
When an object moves along the camera's Z-axis, it produces minimal 2D pixel displacement. Instead, it changes scale. As Hanyu Zhou et al. and Sizhuo Ma point out, standard 2D optical flow algorithms cannot reliably distinguish these scale changes from actual depth-based motion without explicit 3D context, leaving a significant gap in situational awareness.
Cross-Modal Fusion: Pairing Complementary Blind Spots
The most direct way to resolve these individual blind spots is to pair sensors with orthogonal tracking strengths. Because a camera captures tangential motion across its image plane and a radar captures radial motion via frequency shifts, fusing the two modalities allows systems to mathematically recover the full velocity vector.
[Camera Optical Flow] ---> Captures Tangential (X, Y)
+ ===> Full 3D Velocity Vector
[Doppler Radar Return] ---> Captures Radial (Z)
To achieve this, Yunfei Long et al. proposed a framework that pairs radar returns with camera optical flow. By establishing a closed-form mathematical relationship between the 1D radial Doppler signal and 2D pixel velocities, their system resolves the missing directional components to compute full 3D velocity vectors for moving targets.
In a similar vein, Bin Yang et al. developed RadarNet, which fuses the geometric precision of LiDAR with the instantaneous velocity data of radar. By using an attention-based late fusion mechanism, the network leverages historical spatial data to constrain and resolve the unobserved tangential velocity components of radar detections.
While cross-modal fusion is highly effective, it introduces a secondary challenge: training models to understand these relationships without massive manual labeling efforts. To bypass this bottleneck, Fangqiang Ding et al. introduced CMFlow, a framework that uses cross-modal supervision. By using unsupervised signals from surrounding cameras and LiDAR during training, the network teaches a 4D radar model to predict complete 3D scene flow vectors, though the authors note that supervising purely tangential motion remains an ongoing challenge.
Spatiotemporal Triangulation: Changing the Perspective
When multi-sensor arrays are unavailable or cost-prohibitive, researchers turn to methods that resolve ambiguity by changing the perspective of a single modality—either across space or across time.
Spatial Multi-View Triangulation
If a single radar is blind to an object's sideways motion, placing a second radar at a different location changes the line-of-sight vector $\mathbf{\hat{r}}$. What was a purely tangential movement to the first sensor becomes a detectable radial movement to the second.
Soheil Hor et al. leveraged this concept in MVDoppler to classify human gait patterns. Because micro-motions of limbs vanish when a person walks perpendicular to a single radar beam, their system fuses data from multiple radar perspectives. This multi-view architecture ensures that complex motion signatures remain visible regardless of the subject's walking trajectory.
Temporal Triangulation
Alternatively, a single sensor can observe an object over time. As an object moves tangentially across a sensor's field of view, its relative angle to the sensor changes. By tracking the object's changing spatial coordinates across sequential frames, downstream models can infer the unobserved velocity components.
This temporal tracking underpins trajectory prediction frameworks like LiRaNet, which integrates sequential radar returns to resolve the 1D Doppler signal into a reliable 2D motion vector. However, relying purely on time introduces latency. As Fraser Rennie et al. demonstrate, instantaneous state estimation in scanning radars remains heavily biased toward longitudinal motion, meaning that sudden lateral or rotational maneuvers can still cause tracking lag before the temporal model adapts.
Breaking the Far-Field Assumption with Spherical Waves
The most mathematically elegant solution to tangential blindness avoids multi-sensor fusion and temporal tracking entirely. Instead, it rethinks the fundamental physics of how wireless signals propagate using near-field sensing.
Traditional radar and wireless communication systems operate under the far-field assumption. This model assumes that the electromagnetic wavefronts striking an antenna array are perfectly flat, parallel planes. Because the waves are planar, the angle of arrival is identical across every antenna element in the array, yielding a uniform Doppler shift that only reveals a single radial velocity component.
FAR-FIELD MODEL (Planar Waves)
Target (Far Away) ======> | | | | ======> [ Antenna Array ]
Uniform Doppler shift across all elements.
Tangential velocity is invisible.
However, as wireless systems shift toward Extremely Large Aperture Arrays (ELAAs) and higher frequencies (such as mmWave and THz), the target often falls within the sensor's near-field boundary. As explained by Zhaolin Wang et al., near-field electromagnetic wavefronts cannot be treated as flat planes; they are inherently spherical.
NEAR-FIELD MODEL (Spherical Waves)
Target (Close) ======> ) ) ) ) ======> [ Antenna Array ]
Non-uniform Doppler shift across elements.
Tangential velocity becomes readable!
Because the wavefront is spherical, each individual antenna element in the array observes the moving target from a slightly different relative angle. This geometric divergence means that a target moving tangentially will produce a non-uniform Doppler shift across the face of the receiver array.
By measuring this spatial gradient of frequency shifts across a single aperture, the system can mathematically decouple the radial and transverse components of motion. As a result, near-field processing allows a single base station to instantaneously estimate an object's full velocity vector, turning what was once a physical blind spot into a rich source of spatial data.
Key Foundational Work
Separating Non-Stationary from Stationary Scene Components in a Sequence of Real World TV Images (1977) — contributed 1 concept elements to this direction.
Real-time motion compensation for EM bronchoscope tracking with smooth output - ex-vivo validation (2012) — contributed 1 concept elements to this direction.