Week 6: Roo bars, AEB, and their blind spots

What a safety technology actually does, under what conditions it's been shown to work, and what evidence would justify recommending one.

29 March 2027

Open the slides

The Brochure Has Excellent Daytime Vision.

Week 5 ended on an assumed number: a detection distance, fed into a stopping-distance model, with no account of where that number would actually come from. This week asks the question directly. Detection distance isn’t just something a driver’s eyes provide — it’s something a piece of technology can claim to provide too, and that claim deserves the same scrutiny a physics assumption gets.

A polished claim#

“Our system detects wildlife before you do.”

Read that sentence the way you’d read any other unverified claim in this unit: what, exactly, is it saying? Which wildlife — a kangaroo at 40 metres, or a bandicoot at 5? At what distance, and in what lighting — the glossy daylight photo on the box, or the dusk conditions Week 5 spent an entire lecture on? And with what consequence if the detection happens — does the vehicle brake, or does a light on the dashboard turn on while the driver’s attention is elsewhere? A claim that answers none of these questions hasn’t told you anything you can act on yet.

Four different functions#

“Safety technology” bundles four functions that a fair evaluation has to keep separate:

• detection — identifies that a possible hazard is present, using whatever sensor the system relies on (camera, radar, thermal, or some fusion of these);
• warning — alerts the driver that a detection has occurred, typically through a sound, light, or dashboard message;
• automatic emergency braking (AEB) — changes the vehicle’s own motion in response to a detection, without waiting for the driver;
• impact protection — changes some aspect of what happens if a collision occurs, regardless of whether anything was detected beforehand.

A roo bar is squarely in the fourth category, and only the fourth category. It does not detect an animal. It does not brake the vehicle. Its presence on the front of a car changes nothing about whether a collision happens — only, potentially, some aspects of what that collision does to the vehicle if it occurs. Treating “roo bar fitted” and “collision risk reduced” as the same claim conflates a structural fitting with a sensing system, and the specification cards later in this lecture keep that distinction explicit.

False positives, false negatives, and a small worked example#

Any detection system makes two kinds of mistake, and a vendor’s headline accuracy figure can hide either one:

• a false negative — an animal is present and the system fails to detect it;
• a false positive — no animal is present and the system reports one anyway.

Suppose a fictional system is trialled on 100 staged approaches: 40 of them have an animal actually present near the road, and 60 do not. Of the 40 animal-present trials, the system correctly detects 34 and misses 6 — a false negative rate of 6/40 = 15%. Of the 60 no-animal trials, it correctly stays quiet on 51 and falsely alerts on 9 — a false positive rate of 9/60 = 15%. Notice the denominator changes: false negative rate is counted against animal-present trials, false positive rate against animal-absent trials. A single combined “accuracy” figure across all 100 trials can look impressive while hiding a false negative rate that matters far more to a driver than a false positive one does.

Detection distance is a distance margin input, not the whole story#

A stated detection or warning distance plugs directly into Week 5’s distance margin:

Distance margin = assumed detection distance − modelled stopping distance

A vendor’s specification sheet is, in effect, supplying the first term. A 120 m stated detection range at 80 km/h (Week 4/5’s 71.3 m stopping distance) gives a margin of +48.7 m — comfortable, on paper. But if that same system’s own documentation admits its range drops to 60 m once the target is partly obscured by vegetation, the margin at that reduced range is 60 − 71.3 = −11.3 m. The headline range and the fine-print range are not the same claim, and only one of them describes the roadside scrub Week 5 spent a scene card on.

Detection rate is not collision-prevention rate#

A system that detects 85% of animals present is not a system that prevents 85% of collisions. Between detection and prevention sit warning comprehension time, the driver’s own reaction, whether the vehicle is capable of stopping in the distance the warning bought it, and whether the animal stays still long enough for any of that to matter. A high detection percentage is a claim about the sensor. A collision-prevention percentage is a claim about the entire chain from sensor to outcome, and demonstrating one does not demonstrate the other.

Performance varies with conditions#

The same system’s real performance is not one number. It typically varies with:

• lighting — daylight, dusk, full night and headlight glare are different sensing problems for a camera-based system, the same way they’re different problems for a human eye;
• animal type and posture — a stationary, broadside kangaroo and a fast-moving, end-on wallaby present different signatures to most sensors;
• obstruction — roadside vegetation that hides an animal from a driver can hide it from a sensor too, though not necessarily in the same way;
• speed — a system validated at suburban speeds is not automatically validated at highway speeds, and braking physics doesn’t get gentler just because the manufacturer didn’t test that far.

A specification card that states performance under only one of these conditions has told you about one condition.

Designing a fair test#

A test designer who wants a favourable result can get one without lying, simply by choosing conditions: daylight only, one animal size, one speed, a driver’s-eye camera angle instead of a real approach geometry. A fairer test states its conditions in advance, includes the conditions the vendor would rather skip (dusk, obstruction, a range of speeds and animal types), reports a real denominator for every rate it quotes, and is explicit about what it did not test rather than letting silence imply success. Being able to spot which of these a given test skipped is the actual skill this week is teaching.

Specification cards#

Three fictional systems and a roo bar, each specified consistently and honestly, including what’s unknown:

System A — optical wildlife-warning. Detection + warning only; no braking. Tested in daylight, clear weather, straight rural road. Stated range: 90 m detection in daylight; the manufacturer’s own documentation notes reduced performance below approximately 20 lux (dusk and below). Latency: 0.4 s from detection to driver warning. Supported targets: large, stationary-to-slow-moving animals of kangaroo/deer size; not validated for small or fast-moving animals. Known limitation: no independent low-light validation exists for this system.

System B — thermal wildlife-warning. Detection + warning only; no braking. Tested at dusk and night; performance degrades in rain and fog. Stated range: 120 m; the vendor’s own published test note states this drops to approximately 60 m when the target is partially obscured by vegetation. Latency: 0.6 s processing delay. Supported targets: medium and large mammals with a distinguishable heat signature; not validated for reptiles or animals below a stated size threshold. Known limitation: warning only — this system does not brake the vehicle under any condition.

System C — automatic emergency braking (radar/camera fusion). Detection + automatic braking. Tested against a limited target-class set (pedestrian-shaped and vehicle-shaped targets) at speeds up to 60 km/h. Activates on a time-to-collision threshold rather than a fixed distance. Latency: approximately 0.3 s from detection to brake actuation, for the tested target classes only. Supported targets: pedestrian-shaped and vehicle-shaped targets, per the manufacturer’s own test matrix. Wildlife performance is explicitly unverified — no published wildlife trial exists, and no data exists above 60 km/h.

Roo bar — impact protection. Does not detect, warn, or brake. A structural fitting that may change some vehicle-damage outcomes in specific, aligned, frontal-impact scenarios. Evidence is still needed on compatibility across vehicle makes and models, on occupant injury outcomes, on animal injury and mortality outcomes, and on effects on other road users — none of these are claimed, either positively or negatively, by fitting one.

Where this leaves you#

None of the three fictional systems above has been shown to prevent a collision; each has been shown to do something narrower, under stated conditions, with a stated gap in the evidence. That gap is not a reason to dismiss the technology — it’s the reason to state, precisely, what would still need demonstrating before a recommendation could honestly go further than “conditionally promising.” “Wildlife detection” is a category. “The kangaroo behind that bush” is a test case, and a specification sheet that never mentions the bush hasn’t addressed it.

Where this connects to your assessment#

This week’s habits — checking a claim’s denominator, separating detection from prevention, reading a specification’s stated limits rather than its headline — are exactly what the literature review and hotspot analysis asks you to bring to a much larger body of published evidence. That assignment is due 12 April 2027, and is worth 40% of the unit; it is not due this week, and nothing in weeks 4 to 6 adds a new graded requirement to it. Weeks 4 through 6 have been preparation — the physics, the perception, and now the technology critique — for the review you’ll write once you’ve had time to apply all three.

A warning delivered after the encounter is a notification, not an intervention. The brochure has excellent daytime vision.

Reading#

Huijser, M. P., Mosler-Berger, C., Olsson, M., & Strein, M. (2015). “Wildlife Warning Signs and Animal Detection Systems Aimed at Reducing Wildlife-Vehicle Collisions”, in Handbook of Road Ecology (pp. 198–212). Wiley. A review of real detection and warning systems and the evidence for their effectiveness, which varies considerably by system and by study — the same caution this lecture applies to its own fictional specification cards.