01 Pressure waves: what sound is
A loud exhaust pushes on the air around it. That push spreads out as a ripple of high and low pressure — that ripple is the sound.
Only two facts about it matter for this project:
- It moves at a fixed speed: 343 m/s. One metre takes about 3 milliseconds. This is the fact the whole noise camera is built on — it is how the microphones will find the car in section 4.
- It fades fast as it spreads out. You know this already: to talk to someone an arm's length away you speak softly, but to reach a friend across the street you have to shout — and even a shout barely gets there. Loudness drops off with distance much faster than you would guess. That is why the measuring equipment in this project sits close to the road, about 7.5 m from the passing traffic — any further back and the noise would arrive too faint to judge. (Section 3 puts a number on “fast” once we have a unit for loudness.)
Below, the exhaust is the dot. Red is high pressure, blue is low. The dark ring is one wavefront travelling outward at 343 m/s.
02 Frequency and wavelength
Frequency and wavelength are two ways of describing the same ripple:
- Frequency is how fast the air wiggles — the pitch you hear. Fast wiggle = high note (a brake squeal, ~4 kHz); slow wiggle = low note (an exhaust rumble, ~100 Hz).
- Wavelength is how far apart the ripples are, in metres — the same wave seen in space instead of in time.
They are locked together, because every sound travels at the same 343 m/s. A fast wiggle packs its ripples close together (short wavelength); a slow wiggle spreads them far apart (long wavelength):
The sizes come out bigger than you might expect. A low exhaust rumble has ripples about 3.4 m apart — wider than a car. A high tyre squeal has them about 8 cm apart — smaller than your hand.
Here is why that one number decides everything. Wavelength is the finest detail a sound can carry — think of it as the sound's ruler. A ruler marked only every metre cannot tell you where something is to the centimetre; a finely-marked ruler can. A long-wavelength rumble is a coarse ruler, a short-wavelength squeal a fine one — and we are trying to tell apart cars sitting in 3.5 m lanes. Drag the frequency and watch the ripple shrink against the lane:
| Frequency | Wavelength | On the road | Elsewhere in life |
|---|---|---|---|
| 100 Hz | 3.4 m | exhaust rumble | the lowest note of a bass guitar |
| 700 Hz | 0.49 m | engine under load | the middle of a speaking voice |
| 1 kHz | 0.34 m | general traffic | a telephone dial tone |
| 4 kHz | 0.086 m | tyre / brake squeal | a smoke alarm; where hearing is sharpest |
So the sound EPA most wants to fine — a low, loud exhaust — is exactly the one with the coarsest ruler, the hardest to pin to a single car. Section 5 shows that failure happening live.
03 Decibels: the loudness scale
A decibel (dB) is just a compact number for loudness. Real sounds span an enormous range — the loudest you can stand is about a million times stronger than the faintest you can hear — so instead of giant numbers we squash the range with a logarithm. Two things to remember: 0 dB is the faintest sound a human can hear (it means “barely audible”, not silence), and every +10 dB sounds about twice as loud.
Anchor it to sounds you know. Everything this project measures lives in the shaded rows — a vehicle loud enough to fine sits around 85–100 dB at the roadside:
| Level | Sounds like |
|---|---|
| 0 dB | threshold of hearing — near silence |
| 30 dB | a quiet whisper |
| 60 dB | normal conversation |
| 85 dB | heavy city traffic — prolonged exposure damages hearing |
| 100 dB | a motorcycle at close range, or a nightclub |
| 120 dB | a jet taking off nearby — the threshold of pain |
| 140 dB+ | immediate risk of hearing damage |
A motorcycle is not “a bit louder” than conversation — it carries about 10,000× the sound energy. Decibels compress that huge range into small numbers. Three rules of thumb are all you need:
- +10 dB ≈ sounds twice as loud.
- +3 dB = double the energy (barely noticeable to the ear).
- −6 dB = you doubled your distance from the source.
That last rule matters for enforcement: a bike measured at 95 dB from 7.5 m would read 101 dB from half that distance. The measurement distance must be fixed and known, or the number means nothing in court. Try it:
dB(A) is the same decibel, but filtered to match how your ear actually hears. Your ear barely registers deep bass and extreme treble, so dB(A) turns those down in the measurement the same way — the number then reflects how loud a sound feels, not just its raw energy. “90 dB(A)” means “90 dB as a person experiences it.” That is why noise law, EPA's included, is always written in dB(A) — the “A” is simply the name of that ear-matching filter.
(One aside: the heatmaps inside this app use relative dB — the loudest spot is 0 and everything else negative — because for finding the car only the shape of the sound map matters, not the absolute level.)
04 Delay-and-sum: how the SoundCam finds where
You already do this. In a dark room, a friend says your name and you turn straight to them — because the sound hit the near ear a hair before the far ear, and your brain reads that tiny gap as a direction. The SoundCam is the same trick with 56 ears instead of 2.
One microphone tells you only how loud. But the SoundCam has dozens, spread across about a metre, and it uses the timing gaps between them to work out where. Here is the trick.
Because sound travels at a fixed speed, a noise from any one spot reaches each microphone at slightly different moments — its own pattern of tiny delays. The SoundCam plays a guessing game with that. It picks a spot, works out the delay each mic should have if the sound came from there, slides every recording back by that much, and adds them all together.
Why adding? Picture a crowd clapping. All together → one loud crack. At random → a shapeless mush. Sound waves stack the same way: line the recordings up and their peaks reinforce into one big signal; leave them misaligned and peaks fall on troughs and cancel to almost nothing.
So a right guess makes the recordings line up and add up loud; a wrong guess leaves them out of step and they cancel to quiet. Do this for every spot on the road and only the true source lights up — that grid of loud-and-quiet is the heatmap this app lays over the traffic photo. (“Slide back, then add” is literally the method's name: delay-and-sum.)
Try it: drag the orange focus onto the source and watch the ragged traces snap into line as the beam power leaps to its peak.
That is the entire algorithm, and it is only a dozen lines of arithmetic. Everything that follows is about dealing with its one weakness — which is next.
05 The resolution limit: low frequencies smear
Here is the whole problem in a sound you know. A home cinema needs only one subwoofer and you can hide it anywhere — you simply cannot tell where deep bass comes from. But a jingling set of keys or a squeaky hinge? You can point at it instantly. Low notes are fuzzy to locate; high notes are sharp.
Same physics on the road: a tyre squeal is “treble” and easy to pin to one car, while a low exhaust rumble is “bass” and smears across lanes. One formula says how badly:
Bigger wavelength → bigger blur. Plug in the real hardware — a 0.7 m array, 10 m from the road:
| The sound | Spot size | Verdict |
|---|---|---|
| 4 kHz brake squeal | ~1.2 m | inside one lane — that car, no argument |
| 700 Hz loud exhaust | ~7 m | covers both lanes — could be either car |
See it live. Two vehicles sit in adjacent lanes; the map is the real delay-and-sum, recomputed as you drag. Start at 4 kHz, then slide down to 700 Hz and watch two cars melt into one blob:
The enforcement problem in one sentence: the noise EPA most wants to fine — a loud modified exhaust — is low-frequency, and low-frequency sound is exactly the sound the array cannot pin to one lane. More microphones do not help (try it — the blob stays the same size). A physically bigger array does, but a trailer can only be so wide. The way out is to add information the microphones do not have. That is the next section.
06 Sound → road → car: pinning a hotspot to one vehicle
A hotspot in the sound map is not a point — it is a direction. The noise could have come from anywhere along that line.
The fix is one piece of extra knowledge the microphones don't have: the car is on the road, and the road is flat. Follow the direction until it hits the road — now it is a single point. That point tells you the lane, and projecting it into the ANPR camera's photo drops a marker on the exact car whose plate was just read. Acoustics finds the direction; geometry finds the car.
One real-world catch to try below: exhausts sit ~0.5 m above the tarmac, not on it. Aim at the road surface and the point lands about a metre long — on the car behind, in dense traffic. Toggle the height correction and watch the marker walk back onto the right car.
The road↔image mapping is a homography: one small matrix, calibrated once from four surveyed points on the tarmac. Once you have it, every hotspot can be pushed down to the road and back up into the camera automatically.
Note what geometry cannot fix: a lobe genuinely wider than a lane. Drag the frequency down and the ellipse straddles the boundary no matter how perfect the calibration — section 5's limit is physics, not a calibration error. When that happens the honest answer is “ambiguous”, and a system built for evidence has to drop the case rather than guess.
07 The tender's vocabulary
The EPA tender and the SenSen offer are written in acronyms. Almost all of them describe something on this page. Here they are, explained through one event: a loud motorbike rides past the roadside trailer at 2 a.m.
The one sentence the whole bid turns on: “Anyone can measure the noise; SenSen proves who made it.” Measuring loudness needs one microphone. Proving which vehicle was loud needs sections 4, 5 and 6 of this page.
One pass-by, five boxes
The trailer carries four sensors and a computer. As the motorbike passes, each does one job:
| Box | Its one job | On this page |
|---|---|---|
| Acoem Fusion | How loud? A single precision microphone that turns pressure into a legally defensible number in dB(A). | § 3 |
| SoundCam | Which direction? An array of microphones that beamforms a sound map, so the noise can be traced back to a bearing. | § 4–5 |
| Axis P1488-LE | Which plate? An ANPR camera that reads the registration in the dark, at speed. | § 6 |
| SenDISA | Same vehicle? The computer in the trailer. It intersects the bearing with the road plane, picks the lane, and ties the loud moment to the plate that was there at that moment. | § 6 |
| SenBOS | Is it a fine? The cloud site where an EPA officer opens the packaged evidence — audio, sound map, photo, plate — and decides. | — |
The fragile link is the fourth row. The meter says 96 dB(A) at 02:14:07.31; the camera says plate CA41EX at 02:14:07.28. Tying them together is only fair if the sound map can say which lane — and section 5 showed that for a low rumble it often cannot. Closing that gap is the whole game.
The numbers the meter reports
The tender lists LAmax, LAeq, LAE, LA1–99…
They look cryptic, but each is just a different way of collapsing the same wiggly line into one
number. Here is the line for our 2 a.m. motorbike. Ride it past and watch which numbers move:
One experiment tells the whole story: drag the speed from 20 to 120 km/h. The peak (LAmax) does not move at all — the bike still passes just as close — but the total noise dose (SEL) drops by 7 dB, because a fast bike is gone sooner. The same ride is a serious offence under a peak rule and a mild one under an energy rule. That is why the tender makes the meter report the whole family, not one headline number.
| LAmax · LAminthe peak / the quietest instant | The loudest moment of the pass-by is what a resident actually notices, so “loud exhaust” rules are usually written against LAmax. |
| LAeqthe average | The steady level carrying the same total energy as the real fluctuating one. An energy average, so it sits much nearer the peak than the background. |
| LAE · SELthe total dose | The whole event squeezed into one second — the fair way to compare a quick loud pass with a slow moderate one. |
| LA10 · LA90percentiles | The level exceeded 10% / 90% of the time. Over a whole night, LA10 is “the loud traffic” and LA90 is “the hum between vehicles”. (They need a long window — over one 8-second pass-by, LA90 only reaches the quiet tail, as the widget shows.) |
| A / C / Z weightingLAeq · LCeq · LZeq | A = filtered like your ear (the legal default). C = keeps the bass, so it catches thumping exhausts that A-weighting hides. Z = raw and unfiltered. Reporting all three means nobody can claim the filter chose the verdict. |
The words that make it evidence
Measuring is easy; measuring so the number survives a court challenge is the actual product. Each of these terms answers one objection a lawyer would raise about our 2 a.m. motorbike:
| Class 1“was the meter accurate?” | The laboratory-grade accuracy tier for sound meters (IEC 61672, ~±1 dB). Enforcement requires it — which is why the trailer carries an Acoem Fusion instead of a cheap microphone. |
| Type approval · NATA calibration“was it accurate that night?” | A regulator certified this model of meter, and an accredited lab checked this individual unit against a national standard. Without both, 96 dB(A) is data, not evidence. |
| ANPR · OCR“was it really that plate?” | The camera finds the plate (~98%), reads its characters (~99%), and votes across several frames of the same car to close the remaining gap. |
| GPS time-sync <50 ms“same car as the noise?” | Meter and camera stamp their data from the same satellite clock. At 100 km/h a car moves just 1.4 m in 50 ms — so “the noise and the plate were simultaneous” is a measurement, not an assumption. |
| Chain of custody“was the photo edited?” | Every file is cryptographically signed the instant it is captured, so any later change is detectable. |
| Two-layer AI“was it even a vehicle?” | Most triggers are not offences — a truck's compression brake, a slammed door. One filter at the trailer, a second in the cloud; a human only reviews what survives both. |
| IP66 · IK10“will it last outdoors?” | Sealed against dust and hose-down rain, and rated to shrug off a 5 kg impact. Roadside kit gets weathered and vandalised. |
Strip the costume off and it is the same physics as sections 1–6: the meter answers how loud, the array answers from where, the road plane turns that into which lane, the camera answers whose — and the shared clock staples the four together.
08 Recap
The whole page in three sentences. Sound reaches each microphone at slightly different times, and undoing those delays points back at the source — that is the SoundCam. Low rumbles blur wider than a lane, so the direction alone cannot always name the car. Intersecting it with the flat road — and reading the plate that was standing on that spot at that instant — can.
| The idea | On this page |
|---|---|
| Sound is a pressure wave that fades fast with distance | § 1 |
| Pitch, wavelength, and why wave size sets the finest detail | § 2 |
| Decibels and dB(A) | § 3 |
| Delay-and-sum: turning timing gaps into a direction | § 4 |
| Why a low rumble smears across two lanes | § 5 |
| From a direction to “that car, in that lane” | § 6 |
| The words used for all of it in practice | § 7 |
Every widget on this page ran the real delay-and-sum live in your browser — no server, nothing
pre-baked. Add ?selftest=1 to the address to run the built-in checks that confirm the maths
against known values.