Automotive, security, and healthcare deployment

"A demo works once, on my bench, with my data. A deployment works ten million times, in the rain, on a chip that costs three dollars, for seven years, without a lawyer calling."

A Seasoned AI Agent

This section assumes the radar signal chain and representations of Section 44.2 through Section 44.6, the weather-robust fusion argument of Section 44.5, and a working idea of safety cases and regulation from Chapter 68 and Chapter 70. Our concern here is not new algorithms but the deployment envelope: which band you may transmit in, what evidence a certifier demands, and how a fleet of radars behaves once thousands of them share a street or a hospital ward.

Automotive: spectrum, safety cases, and the interference wall

Automotive radar occupies the 76 to 81 GHz band by international allocation, with long-range cruise and emergency-braking sensors using a narrow slice around 76 to 77 GHz and short-range corner sensors sweeping the wider 77 to 81 GHz for finer range resolution. That choice is not the engineer's: it is fixed by ITU, FCC Part 95, and ETSI EN 302 264, which cap effective isotropic radiated power and occupied bandwidth. A model that wants more bandwidth for sharper range (recall \(\Delta R = c/2B\) from Section 44.1) cannot simply take it; the regulatory mask is a hard boundary on what perception is physically allowed.

On top of spectrum sits functional safety. A radar that arms automatic emergency braking is an item in an ISO 26262 safety case, assigned an Automotive Safety Integrity Level (ASIL) up to D, the most stringent. Two failure modes matter and pull in opposite directions: a missed detection (the radar fails to see a stopped truck) and a false positive (phantom braking on a manhole cover or an overhead gantry). ISO 21448 (SOTIF, safety of the intended function) governs the second, because phantom braking is not a component fault but a limitation of a perfectly working sensor misreading the world. This is why deployed automotive radar almost never acts alone: it is decomposed across redundant channels with camera and, on higher tiers, lidar, exactly the missing-modality-robust fusion argument of Chapter 50, so that no single sensor failure crosses the ASIL threshold.

The constraint that scales with success, and that few benchmarks capture, is mutual interference. Every car radiating an FMCW chirp is noise to every other car's receiver. As market penetration climbs, the probability that two radars in view collide in time and frequency rises, raising the noise floor and manufacturing ghost targets. The rough model below treats each of \(N\) nearby radars as occupying a fraction \(\rho\) of the shared time-frequency resource and asks for the chance that at least one overlaps your own frame.

import numpy as np

def interference_prob(n_radars, duty=0.02, seed=0):
    """P(at least one of n_radars collides with ego in a shared time-freq slot).
    duty = fraction of the shared resource a single radar's chirp train occupies."""
    p_one_clear = (1.0 - duty)          # one interferer misses the ego slot
    p_all_clear = p_one_clear ** n_radars
    return 1.0 - p_all_clear

for n in [1, 5, 20, 50, 100]:
    print(f"{n:4d} interferers -> P(collision) = {interference_prob(n):.3f}")
# 1 -> 0.020 | 5 -> 0.096 | 20 -> 0.332 | 50 -> 0.636 | 100 -> 0.867

The numbers above are why production radars now ship interference mitigation as a first-class feature: chirp-timing randomization, digital detection-and-repair of corrupted chirps, and the emerging IEEE and ETSI coordination proposals. A deployment plan that ignores fleet density passes every solo bench test and then degrades in exactly the dense urban scenes where braking matters most, a distribution shift in the sense of Chapter 66 that you cannot reproduce with a single sensor in a lab.

Security and smart spaces: false-alarm economics and the privacy dividend

In perimeter security, presence sensing, and building automation, radar's selling point flips from safety to privacy plus all-weather reliability. A 60 GHz radar in a ceiling never forms an image, so it can watch a bathroom, a bedroom, or a prison corridor where a camera is legally or ethically impossible, while still counting occupants, detecting intrusion, and measuring dwell. The dominant deployment metric here is not accuracy but the false-alarm rate. A perimeter radar that cries wolf on windblown foliage, rain, or a wandering deer trains its human operators to ignore it, which is worse than no sensor. The detect-then-classify micro-Doppler cascade of Section 44.3 is the standard defense: detect the range-Doppler peak, then classify its motion signature as human versus clutter before raising an alarm.

Security radar also lives under spectrum rules, but friendlier ones. Most indoor presence and gesture radars use the 60 GHz ISM band (57 to 64 GHz), unlicensed under FCC Part 15 and ETSI EN 305 550, which is why consumer devices such as presence-aware thermostats and the gesture radar in some phones could ship at all. The tradeoff is short range and shared spectrum with Wi-Fi-class devices, tolerable for a room, unacceptable for a highway.

Healthcare: the regulatory pathway is the product

Contactless health radar, the vital-sign and sleep and fall sensing of Section 44.6 and Chapter 33, faces the steepest deployment gate in this section: it makes claims about a person's body. The moment a radar reports a respiration rate or flags a fall as a health event, it may be a regulated medical device. In the United States that typically means an FDA 510(k) clearance for a Class II device (demonstrating substantial equivalence to a cleared predicate) or a De Novo request when no predicate exists; in Europe it means the Medical Device Regulation and a notified-body audit. The evidence bar is a prospective clinical study, agreement metrics against a reference monitor (a nasal cannula for breathing, a certified bed sensor or witnessed observation for falls), and, crucially, subject-independent validation, the leakage-safe protocol of Chapter 5 raised to a legal requirement. The full clinical-validation and biometric-privacy playbook lives in Chapter 34; the point here is that for a health radar, that pathway is the deployment plan, and it dwarfs the model-training budget.

A telling deployment reality is the wellness-versus-medical line. Vendors routinely ship the same radar hardware as a general-wellness sleep-quality gadget (no clearance needed, no diagnostic claim allowed) and, on a separate, slower track, as a cleared device that a clinician may act on. The model may be identical; the claim, the labeling, and the evidence are not. Deploying into a hospital adds electrical-safety (IEC 60601) and cybersecurity obligations that a consumer sleep tracker never touches.

What deployment adds on top of the model

Across all three domains a common deployment checklist emerges that no accuracy metric captures. Edge compute: the model must run on the radar's companion microcontroller or a small SoC within a milliwatt and millisecond budget, the quantization-and-pruning discipline of Chapter 59. Calibration drift and updates: a seven-year automotive lifetime or a multi-year hospital deployment needs over-the-air model updates and per-site recalibration, governed by the fleet-ops practices of Chapter 69. Adversarial and spoofing resistance: radar is spoofable with a repeater that injects false targets, a threat treated in Chapter 68. The pattern to internalize is that the model is perhaps a quarter of the deployed system; spectrum compliance, the safety or regulatory case, fleet operations, and privacy governance are the rest, and they decide whether the product ships at all.

What's Next

In Chapter 45, we swap the radio wave for the infrared photon and study thermal and multispectral sensing, another modality that sees through darkness and, like radar, offers a privacy-preserving path to human monitoring, with its own calibration physics, domain-adaptation tricks, and deployment rulebook.