Catching the Moment Frames Miss with an Event Camera

EVENT CAMERA

A 200 fps camera misses 70 % of 0.5 ms events, while an event camera records the same event as a few hundred events.

Taking a bouncing solder ball, I calculated the odds a frame camera misses the event, the speed an event pixel can follow, when flickering light floods the sensor, and how long the time window should be.

About 15 min read

A solder ball thrown up during reflow or selective soldering is a tiny mirror bead tens of μm across. Because a sphere’s surface faces every direction, light from any side leaves one bright point of specular reflection¹ toward the camera, and plated connector pins glint the same way. The solder resist beneath is a matte surface that scatters light evenly, so without motion its brightness does not change. What the inspection is after is the moment, from a few hundred μs to a few ms, when that shiny bead bounces.

A frame camera gathers light while the shutter is open to make one image and records nothing until the next. If an event passes in that gap, raising the frame rate cannot bring back what was never recorded. An event camera² is a sensor whose pixels send a signal only at the moment brightness changes, so it has no such gap, yet it does not see every moment for free either. I weigh the technology with three numbers: the odds a frame misses the event, the speed an event pixel can follow, and the condition under which flickering light floods the sensor.

The odds a frame misses a 0.5 ms event

Assume the event is visible for only 0.5 ms and the camera shoots at 200 fps with a 1 ms exposure. Of each 5 ms frame interval, the stretch in which the event lands in at least one image is just the event length plus the exposure, 1.5 ms, so the odds of capture are 1.5 ÷ 5 = 30 %, and the other 70 % never reach the record. Even a lucky capture has a problem. A 25 μm ball moving at 0.2 m/s travels 200 μm during a 1 ms exposure, so at 8.33 μm per pixel a 3 px ball smears to 24 px through motion blur³. What remains is a streak with the ball’s brightness diluted to an eighth.

Holding the smear within one pixel means cutting the exposure to 8.33 μm ÷ 0.2 m/s = 41.7 μs, and getting an image just as bright means lighting 24 times brighter than for a 1 ms exposure. Never missing a 0.5 ms event with that short exposure requires a frame interval under 0.54 ms, or 1,846 frames per second, which for 1280 × 720 px 8-bit images pours out 1.70 GB every second. Most of it is background where nothing happens. Closing the gap with a frame camera makes the cost grow on two fronts at once, lighting and data.

How fast a change-only pixel can follow

An event camera pixel does not gather light. Each pixel watches the logarithm of brightness, and when the change since its last event crosses a preset contrast threshold⁴, it sends out one event meaning brighter or darker, along with a timestamp. The IMX636 sensor described in Prophesee’s documentation and product brief has 1280 × 720 px with 4.86 μm pixels, a dynamic range⁵ beyond 120 dB and a peak output of 1 Geps (one billion events per second), with a latency⁶ listed as 100 μs or less at 1000 lx.

Sending only changes means data appears only in proportion to the event. When a 25 μm ball travels 100 μm at 0.2 m/s over 0.5 ms, it sweeps 3 × 12 = 36 pixels, and assuming a 4× brightness ratio between ball and background and a contrast threshold of 0.2, each pixel fires 7 events on brightening and 7 on darkening, about 500 in all. Over the same 0.5 ms, the 1,846 frames-per-second camera sends about 850,000 bytes, and 500 events are far smaller than that however many bytes each event takes.

Event pixels still take time to respond. The time the ball takes to cover and pass one pixel is its diameter divided by its speed, 25 μm ÷ 0.2 m/s = 125 μs, a little longer than the 100 μs latency at 1000 lx. My rule is that this dwell time must exceed the latency for events to come out on time and evenly, so under these conditions the speed it can follow is around 25 μm ÷ 100 μs = 0.25 m/s. For faster events, I would first look at brighter lighting to cut the latency, and measure the low-light latency, left blank in the product brief, on samples.

Events from shiny surfaces and flickering light

Because an event camera reacts to changes in brightness, even an object that does not move produces events when the light changes. The first suspect is the lighting. Under an LED switched on and off at 1 kHz by PWM dimming⁷, every lit pixel fires brightening and darkening events at least 1,000 times a second each. If all 1280 × 720 px are lit, that is at least 1.84 billion events per second, beyond the sensor’s 1 Geps peak output, and even with only 30 % of the field lit it is 550 million events, more than half the peak. The 500 events of a bouncing ball get lost in that flood. Limiting the event rate on the sensor side discards the overflow, and nothing guarantees that the discarded events leave out the ball’s.

That is why constant-current drive⁸ is the baseline for event camera lighting. If supply ripple makes the brightness swing by 5 %, the log change is 0.049, smaller than the 0.2 contrast threshold, so no events appear, but the margin shrinks the moment the threshold is lowered to raise sensitivity. The product brief lists anti-flicker, an event filter and an event rate function among the IMX636’s event signal processing functions, and Prophesee’s US patent US11856290B2 (priority date January 26, 2018; granted December 26, 2023) claims a method that finds a frequency pattern in the light a pixel receives and extracts scene information from it. A periodically flickering source can be picked out by its frequency.

Even with stable lighting, shiny surfaces stay tricky. On near-mirror surfaces like plated pins or solder balls, the brightness gradient piles up at the edge of each highlight, so a vibration of just a few μm sharply changes the brightness of the edge pixels and events pour out. On matte resist the same vibration produces almost no events. How many events vibration unrelated to defects produces on shiny surfaces depends on surface gloss and machine vibration, and it cannot be guaranteed before a sample test.

Match the time window to the event

Single events carry little meaning, so the decision bundles the events gathered within a set time window⁹. The longer the window, the more noise events pile up in proportion, so if an event produces 500 events over 0.5 ms, a window of around 0.5 ms fits. Stretch the window to 5 ms and the event still yields 500 events while the background noise grows tenfold. The figure below uses 1 to 5 ms windows as its example, a length suited to slower events that unfold over several ms.

An inspection that watches vibration adds one more criterion. If reading how position changes within a single cycle requires the window to stay under a tenth of the period, a 1 ms window suits vibrations only up to 100 Hz and a 5 ms window only up to 20 Hz. Watching a pin vibrating at 1 kHz takes a window narrowed to 0.1 ms, and fewer events gather in each window, making the decision shakier. Widen the window and the motion within a cycle overlaps in one bundle; narrow it and events run short, so the right value lies in between. I would calculate the window length from the event length and vibration frequency first, then check on samples that enough events gather in that window.

Narrowing the area to judge is another way. Sony Advanced Visual Sensing’s US patent US11825218B2 (priority date July 1, 2019; granted November 21, 2023) covers using regions of interest¹⁰ and regions of non-interest to isolate change events and reduce data from uninteresting events, and its description says the region typically follows the activity. In a process where the spots a solder ball can bounce to are known, opening only those spots filters out many of the events from shiny pins from the start.

Evidence for measuring without stopping

Two papers show what event cameras have done in real inspection. A countersink inspection study posted to arXiv in April 2023 (Salah et al., arXiv:2304.04108) measured countersink depth while moving, unlike conventional methods where the robot stops at each hole, and reports inspection about 10 times faster with a depth precision of 0.025 mm.

A paper by Zhu et al., published online in Advanced Science on September 23, 2025, looked for stains and small scratches on semiconductor chips moving at 1 to 9 mm per second, and tested while varying illumination from 0.145 lx to 1,638 lx, a span of about 11,000 times. The abstract highlights event processing sampling times spanning 300-fold in fast motion and dynamic ranges exceeding 10,000-fold, and the body compares about 120 dB for event mode with 40 dB for frame mode. Both results come from each study’s own samples and lighting, so I read them only as evidence of what is possible and measure our own line separately.

Following a 25 μm ball, and the part frames still play

Before settling on a camera, work out three things on one sheet: how many pixels the ball occupies, whether the pixels can follow its speed, and whether the lighting creates events of its own. Taking 8.33 μm per pixel so a 25 μm ball spans 3 px, the IMX636’s 1280 × 720 px gives a field of view of 10.7 × 6.0 mm at a magnification of 0.58×. To make only the ball stand out bright, a dark-field¹¹ ring light that skims the board at a low angle runs on constant current, and that ring needs room between the lens and the board. With a 35 mm lens plus about 20 mm of extension tube, the thin-lens approximation puts the board about 95 mm from the principal plane of the lens; the actual WD¹² is measured from the front of the lens and comes out shorter, so it has to be checked against the datasheet.

Item	① Minimum defect size	② Optical setup · lighting · WD	③ Algorithm parameters
Resolving the ball	25 μm diameter = 3 px (8.33 μm/px)	IMX636 1280 × 720 px, 4.86 μm pixels, 0.58× magnification, field of view 10.7 × 6.0 mm. 35 mm lens + about 20 mm extension tube, object distance about 95 mm (thin-lens approximation; check WD from the lens front on the datasheet)	Ball dwell per pixel 25 μm ÷ speed ≥ latency 100 μs (1000 lx). 125 μs at 0.2 m/s
Lighting · reflection	Ball highlight contrast 4× (assumed)	Low-angle dark-field ring light on constant current (no PWM dimming). Check the WD left after the ring height	PWM at 1 kHz gives at least 1.84 billion events/s over all pixels > 1 Geps. A 5 % ripple is 0.049 in log < threshold 0.2
Time window · region	0.5 ms event, about 500 events	Open only the spots where balls can bounce as the region of interest	Time window around 0.5 ms. For vibration, window ≤ a tenth of the period (a 1 ms window covers up to 100 Hz)

Only when the three rows mesh does an event camera work as equipment that catches events between frames. Lighting brightness sets the speed the pixels can follow, and keeping that lighting from creating events of its own is shared among the drive method, the time window and the region of interest.

There are also clear places where frame cameras fit better. In work that needs one complete image of a moment, such as measuring a stationary part or judging color and texture, an event camera, which sends nothing about places where nothing changes, is at a disadvantage. If events last several ms or longer and move slowly, a 200 fps frame camera with a short exposure may be enough, and setups that combine the two are possible too. Which side wins on a real line depends on event length and surface reflectance, so it cannot be guaranteed before a sample test.

Working out the gaps between frames in numbers is the first job for an engineer choosing how to inspect a high-speed process, and the same question comes back the moment you take on process monitoring. If you get an event camera quote, it is safer to have the lighting drive method and measured low-light latency in the same table. This article explains technology and is not investment advice. See our Disclaimer.

Field Note

If a solder ball bouncing between two frames had to be caught, my first measurement would be how many ms the event stays visible and how fast the ball moves, using a few seconds of high-speed camera footage. If the time the ball covers one pixel comes out longer than 100 μs, the event camera goes on the candidate list, and the event rate with nothing moving under the constant-current ring light gets logged as the floor. If that floor is large enough to bury the 500 events of a single bounce, the lighting and region of interest get fixed before the time window is set.

Field Checkpoints

Once the event length and the floor event rate are in hand, the list below settles whether an event camera comes in.

• WD: with the 35 mm lens and extension tube, does the actual WD from the front of the lens to the board leave room after subtracting the height of the low-angle ring light? (thin-lens object distance about 95 mm)
• Is the time the ball covers one pixel (ball diameter ÷ speed) longer than the latency under the actual lighting? (100 μs or less at 1000 lx)
• Is the lighting on constant current, with no PWM-dimmed source anywhere in the field of view?
• Has the floor event rate with nothing moving been logged?
• Was the time window set from the event length and vibration frequency (window ≤ a tenth of the period)?
• Are only the spots where balls can bounce open as the region of interest?
• Have events from vibrating shiny pins been told apart from ball events on samples?

Frequently Asked Questions

Can existing vision algorithms run on event camera output as is?

The output is a list of per-pixel changes in place of images, so it is hard to use as is. Events can be stacked per time window into something image-like and fed to existing algorithms, but the result shifts with the window length and part of the timing information in the events is thrown away. I would put the effort of building a new processing pipeline into the adoption quote from the start.

What does a 120 dB dynamic range mean for inspection?

A 120 dB range is a brightness ratio of one million. It means there is plenty of room for pixels to avoid saturating or sinking even with a shiny ball’s highlight and dark resist in the same field, which eases the frame-camera problem of losing dark areas when the exposure is cut to spare the highlights. A wide dynamic range does not make flicker or vibration events go away, though, so the lighting drive and time window design are still needed.

What about a setup that pairs it with a frame camera?

The work can be split, with the event camera catching the event and a frame camera keeping a still image of where it happened. Using a burst of events as the signal to fire the frame camera keeps data small while preserving a record. Aligning the two cameras in field of view and timing needs its own calibration, and whether that precision is enough for the decision gets checked on line samples.

Glossary

• [1] Specular reflection specular reflection
  Reflection that leaves a smooth surface in one direction at the same angle the light came in. It is strong on solder balls and plated pins. ↩
• [2] Event camera dynamic vision sensor
  A sensor whose pixels send signals asynchronously only when a brightness change crosses a threshold. Also called a neuromorphic vision sensor. ↩
• [3] Motion blur motion blur
  The smear that appears when an object moves during the exposure. The smear length is speed multiplied by exposure time. ↩
• [4] Contrast threshold contrast threshold
  The size of log-brightness change needed to fire an event. Lowering it raises sensitivity and noise events alike. ↩
• [5] Dynamic range dynamic range
  The ratio between the brightest and darkest light a sensor can handle at once. 120 dB is a factor of one million. ↩
• [6] Latency latency
  The time from a brightness change to the resulting event. It tends to grow as light dims, so it is read together with the illuminance. ↩
• [7] PWM dimming PWM dimming
  Adjusting LED brightness by switching it on and off rapidly and varying the on-time ratio. An event camera sees it as flicker. ↩
• [8] Constant-current drive constant-current drive
  A drive method that keeps a steady current through the LED to hold its brightness. It is the basic guard against flicker events. ↩
• [9] Time window time window
  The time span over which events are bundled for a decision. Too long piles up noise; too short leaves too few events. ↩
• [10] Region of interest ROI
  The area from which the sensor or processing accepts events. Events from the rest are dropped to cut data. ↩
• [11] Dark-field illumination dark-field lighting
  Lighting at a low angle that makes flat surfaces look dark and only raised or rounded objects look bright. ↩
• [12] WD working distance
  The distance from the front of the lens to the inspected part. It decides how much room fixtures like a ring light have in between. ↩

References

• Prophesee, “IMX636 HD Sensor” documentation (1280 × 720 px, 4.86 μm pixels, HDR beyond 120 dB, 1 Geps peak). docs.prophesee.ai
• Prophesee, “IMX636 Product Brief 2024 v3.0” (latency 100 μs or less at 1000 lx; anti-flicker, event filter and event rate functions; pixel ROI). PDF
• Prophesee SA, US11856290B2 “Method and apparatus of processing a signal from an event-based sensor”, priority date 2018-01-26, granted 2023-12-26. Google Patents
• Sony Advanced Visual Sensing AG, US11825218B2 “Dynamic region of interest (ROI) for event-based vision sensors”, priority date 2019-07-01, granted 2023-11-21. Google Patents
• M. Salah et al., “High Speed Neuromorphic Vision-Based Inspection of Countersinks in Automated Manufacturing Processes”, arXiv:2304.04108, 2023-04-08. arXiv
• S. Zhu et al., “Ultrafast Dynamic Defect Inspection With Computational Neuromorphic Imaging”, Advanced Science 12, e10338, published online 2025-09-23. PDF

Related reading — Keeping Frame Rate and Lighting from Colliding on a High-Speed Inspection Line · Moving Parts Are Frozen by How Long the Light Stays On · A Board at 0.5 m/s Leans 15 mm Under a Rolling Shutter