Cyclist Detection
How 77 GHz Radar Classifies a Bicyclist
Cyclist detection sits at the difficult middle of the vulnerable-road-user problem. A bicyclist is faster and larger than a pedestrian yet far smaller and more deformable than a vehicle, and the most dangerous encounters happen at intersections where the bicycle crosses the host path with little or no radial velocity. The radar front end transmits a frequency-modulated continuous-wave (FMCW) chirp sequence and recovers range from beat frequency, range-rate from chirp-to-chirp phase progression, and azimuth from the phase gradient across a multiple-input multiple-output (MIMO) virtual array. None of these primary measurements alone distinguishes a cyclist; the discriminating information lives in the fine Doppler structure.
The defining feature is the micro-Doppler signature. A rigid target produces a single Doppler line at its bulk velocity, but a cyclist is an articulated, multi-part scatterer. The rider's legs pedal at a cadence near 1 to 1.5 Hz, the upper rim of each wheel momentarily moves at twice the bicycle ground speed while the lower rim approaches zero, and the frame and torso contribute the bulk return. A time-frequency spectrogram of the received signal therefore shows a central velocity line flanked by periodic sidebands and high-velocity wheel-rim spikes. This pattern is distinct from the symmetric, lower-amplitude limb modulation of a walking pedestrian and from the clean single line of a car, which is what makes reliable classification possible.
Because crossing cyclists can fall into the zero-Doppler clutter bin, detection cannot rely on moving-target indication alone. Systems extend the coherent processing interval to 20 to 40 ms to resolve the weak micro-Doppler sidebands, apply Doppler-division or time-division MIMO to preserve angle accuracy, and fuse the radar track with a camera classifier so that a tentative radar detection with the right kinematics is confirmed visually before an autonomous emergency braking command is issued.
Angular Resolution and Aperture
Separating a cyclist from an adjacent vehicle or guardrail is fundamentally an angular-resolution problem. The achievable azimuth resolution scales with wavelength divided by aperture, so at the 3.9 mm wavelength of 77 GHz a designer must synthesize a large MIMO virtual aperture to resolve a bicycle 0.5 m beside a car at 30 m. A modest 3-transmit by 4-receive array yields only a 12-element virtual aperture of about 21 mm, which gives roughly 10 degrees of azimuth resolution, too coarse for that geometry. Resolving 0.5 m at 30 m needs near 1 degree, which calls for a virtual aperture around 220 mm: in practice a cascaded multi-chip imaging radar that synthesizes on the order of a hundred virtual elements, fine enough to keep the cyclist track separate through an intersection crossing.
Governing Relationships
fD = 2 × vr / λ, λ = c / f0 ≈ 3.9 mm at 77 GHz
Wheel-rim peak velocity (relative to ground):
vrim,max ≈ 2 × vbike (top of wheel), vrim,min ≈ 0 (contact point)
Azimuth angular resolution:
θres ≈ λ / Daperture (radians)
Where vr = radial velocity, λ = wavelength, c = 3 × 108 m/s, f0 = carrier frequency, Daperture = MIMO virtual aperture. Example: a cyclist at vbike = 6 m/s gives fD ≈ 3.1 kHz, with wheel-rim returns reaching ≈ 6.2 kHz; a 220 mm virtual aperture yields θres ≈ 1°.
Vulnerable Road User Comparison
| Target Class | Typical Speed | Radar Cross Section | Micro-Doppler Cue | Detection Challenge |
|---|---|---|---|---|
| Cyclist | 4 to 8 m/s | -3 to +3 dBsm | Pedal cadence + wheel-rim spikes | Zero-Doppler at crossings |
| Pedestrian | 1 to 2 m/s | -10 to -5 dBsm | Symmetric limb swing | Low RCS, low velocity |
| Motorcycle | 8 to 30 m/s | 0 to +5 dBsm | Strong bulk line, weak micro-Doppler | Narrow frontal profile |
| Passenger car | 0 to 35 m/s | +10 to +20 dBsm | Single clean Doppler line | Easy (reference class) |
| Roadside clutter | 0 m/s | Varies | None (stationary) | Ghosts via multipath |
Frequently Asked Questions
How does a 77 GHz radar tell a cyclist apart from a pedestrian?
It compares micro-Doppler signature and bulk velocity. A pedestrian walks at 1 to 2 m/s with symmetric limb modulation and low RCS (-10 to -5 dBsm). A cyclist moves at 4 to 8 m/s, shows a periodic pedal cadence near 1 to 1.5 Hz plus high-velocity wheel-rim spectral lines, and presents a larger, steadier RCS (-3 to +3 dBsm). A classifier fuses range, range-rate, RCS, and the micro-Doppler spectrogram to separate the classes with greater than 95 percent reliability above an adequate SNR margin.
What angular resolution does a radar need to separate a cyclist from a nearby vehicle?
Resolution θres ≈ λ / Daperture. At 77 GHz (λ = 3.9 mm) a 50 mm virtual array gives about 4.5°, but resolving a cyclist 0.5 m beside a car at 30 m needs near 1°, requiring a roughly 220 mm MIMO virtual aperture. A basic 3-transmit by 4-receive array spans only about 21 mm (roughly 10°), so cyclist-capable systems use large cascaded imaging arrays with on the order of a hundred virtual elements; too little resolution merges the cyclist with an adjacent vehicle or guardrail.
Why is cyclist detection harder at an intersection than on a highway?
A crossing cyclist has near-zero radial velocity, so its bulk return lands in the stationary-clutter Doppler bin and is suppressed by moving-target indication. Detection then leans on weak leg-and-wheel micro-Doppler, which needs 20 to 40 ms coherent processing to resolve, while building and parked-car multipath adds ghost targets and angle errors. These crossing scenarios feature heavily in Euro NCAP cyclist AEB protocols.