Free fall optical sorting: align views and ejector timing

A free-fall sorter must connect three measurements: what each camera sees, which object those observations describe, and when that object reaches an ejector. Opposing cameras improve visibility, but they do not automatically provide complete surface coverage or reliable object matching. This guide is for teams specifying the imaging and timing chain for small produce. Unlike conveyor camera selection, free-fall design requires a velocity estimate independent of belt motion and a clear treatment of acceleration.

Define the inspection geometry and coverage limits

Start with a thin, well-separated curtain of material. Define the inspection width, stream depth, expected object sizes, and distance from the imaging plane to the effective rejection plane. Measure that last distance along the object trajectory, not to the valve body.

Place opposing line scan cameras so their inspection lines observe approximately the same crossing plane. Each camera measures a narrow strip across the curtain. Successive strips form an image as objects pass through that plane.

Two opposing views are not a guarantee of complete surface inspection. Curved surfaces become strongly foreshortened near their visible edges. Recesses, contact between objects, and overlapping trajectories can hide defects from both views. Rotation may reveal additional surface, but uncontrolled rotation cannot be counted as verified coverage.

Define acceptance in terms of defects and observable regions. For example, a requirement to detect a contrasting patch on an exposed face is different from detecting every recessed blemish anywhere on the surface.

Use the field-of-view calculation guide to establish width and sampling. Then check image-circle coverage and edge performance using the line-scan lens selection guide.

Choose spatial sampling before choosing line rate

Select the smallest relevant feature and the number of samples needed across it. A sampling calculation establishes geometric feasibility; it does not establish reliable defect detection without adequate contrast and optical resolution.

For an illustrative inspection width of 400 mm and 4096 samples across the stream:

These are application requirements, not operating claims for any product shown below. Check the complete camera configuration against the required rate.

A conventional conveyor arrangement illustrates the camera, line illumination, and motion reference. In free fall, however, a feeder or conveyor encoder does not directly measure the position of an airborne object.

Use the calculator with the local crossing-speed estimate, then repeat the calculation at the expected speed limits.

Exposure also limits useful resolution. At 5 m/s, an exposure of 10 microseconds produces approximately 0.05 mm of motion during integration:

Exposure shorter than the line period does not by itself guarantee useful images. Confirm that illumination provides sufficient signal without excessive gain. The line-rate setup procedure covers the acquisition settings after these sampling requirements are established.

Estimate velocity at the inspection plane

A falling object accelerates, but a fixed inspection plane does not necessarily require a continuously increasing line rate. With a repeatable release height and release velocity, successive objects can cross that plane at similar speeds.

Velocity variation arises from changing release conditions, bounce, aerodynamic effects, collisions, and different trajectories. An individual object also accelerates while its silhouette passes through the inspection line. A constant-rate image therefore has a time axis that is only approximately a distance axis.

For an initial estimate, neglect aerodynamic drag and collisions:

Validate that estimate using measurements on representative material. Two vertically separated optical observations can estimate transit speed. Their separation, timestamp accuracy, and association reliability determine whether that estimate is useful.

For a variable stream, choose between fixed-rate acquisition with later geometric correction and variable-rate triggering supported by the acquisition hardware. Changing the trigger rate can improve sampling uniformity, but it complicates exposure limits, buffering, and cross-view registration.

Color cameras may also acquire their color channels on spatially separated sensor rows. Where applicable, confirm the documented registration method and test it across the velocity range. Acceleration or rotation can produce colored edges that resemble defects even when the raw channels are individually sharp.

Synchronize observations, then associate objects

A common start signal is not sufficient evidence that two cameras observed the same event simultaneously. Establish what each timestamp represents: exposure start, exposure midpoint, line delivery, or host arrival. Host arrival time includes transport and buffering delays.

Prefer an acquisition arrangement with a documented shared timing reference. Measure the residual offset using an event visible in both views. At 5 m/s, a timing mismatch of 20 microseconds corresponds to 0.10 mm of motion.

Calibrate each view into common coordinates. Opposing views normally need a lateral-coordinate reversal as part of that transformation. Lens distortion and perspective also matter. Calibration at one depth does not eliminate parallax throughout a thick stream.

Use the calibration target coverage guide when planning measurements across the inspection width.

Do not force every observation into a pair. Touching objects, occlusion, or two similar objects crossing together can make association ambiguous. Record unmatched and multiply matched observations separately. Define whether those cases are accepted, rejected, or diverted according to the process requirements and collateral-rejection risk.

Predict arrival from a physical reference point

Every tracked object needs a consistent reference point. A leading edge, estimated center, and trailing edge produce different timestamps. Choose a reference suitable for the ejector footprint, and preserve its definition through acquisition, segmentation, and tracking.

The simplest ballistic estimate starts with downward speed at that reference event and distance remaining to the rejection plane:

Here, h is the remaining vertical drop and v0 is the downward velocity at the imaging reference. This model assumes vertical motion with negligible drag and no intervening collisions.

For an illustrative speed of 4 m/s and a remaining drop of 0.30 m:

That difference represents roughly 28 mm of additional travel near the rejection plane. Ignoring acceleration can therefore move the effective rejection window substantially away from the intended target.

Real produce may depart from the ballistic model. Validate arrival predictions by observing passage at the rejection plane across object sizes and operating conditions. If residual errors correlate with object size, trajectory, or speed, improve the estimator rather than adding one universal timing offset.

Schedule ejector commands against absolute deadlines

Keep the acquisition timestamp with the object record. Processing completion is not a replacement for the time at which the object was observed.

Measure the delay from an electrical command to useful rejection force at the object path. Include output transport, valve response, and jet establishment where relevant. Separate that physical delay from image transfer and classification latency.

Using the preceding flight estimate, suppose a classification decision becomes available 8 ms after the reference event. Assume a measured output delay of 1 ms and actuation delay of 3 ms:

These are illustrative assumptions, not component specifications. Processing latency has already elapsed when the decision becomes available; do not subtract it twice from the absolute command timestamp.

A deadline-aware output queue should detect late decisions explicitly. It should also identify conflicting commands, actuator recovery limits, and pulses that would affect neighboring objects. Lateral drift between inspection and rejection must be included when selecting an ejector lane.

Keep process rejection logic separate from machine safety functions. Validate guards, pneumatic isolation, and safe operating procedures through the machine's safety design.

Size the acquisition path for sustained load

Calculate payload before selecting an interface or host acquisition path. For an illustrative image stream with 4096 samples per line, 50,000 lines/s, and three delivered bytes per sample:

This excludes transport overhead, metadata, alignment, and additional memory copies. A GigE connection cannot carry this uncompressed example. Reduce the requirement or select a verified higher-bandwidth configuration; do not assume the catalog examples below meet it.

At that combined payload, retaining 100 ms of raw image data requires approximately 123 MB before additional processing buffers. Buffer capacity absorbs temporary delays but cannot fix a sustained throughput deficit.

Select frame grabbers and host interfaces using the actual pixel format, camera configuration, and simultaneous acquisition requirements. The Camera Link compatibility guide explains why connector and interface names alone do not establish a working pairing.

Check acquisition behavior independently of classification using the frame-grabber acquisition guide. Then add processing and logging while monitoring dropped lines, queue age, and deadline misses.

Objects can straddle software image boundaries. Preserve object state across those boundaries using the principles in the line-scan overlap guide, without counting an overlap region as another physical object.

Light each view for the classification task

Choose line scan lights for contrast at the required exposure, not simply for an illuminated length. Color grading needs repeatable channel response and stable backgrounds. Shape or foreign-material detection may need a different illumination direction or spectral choice.

Opposing views create an additional problem: illumination intended for one view can enter the other camera or change its background. Test each light independently, then both together. Adjust angles, backgrounds, and optical shielding while checking the complete inspection width.

A suitable physical length does not prove sufficient irradiance at the object. The line-light exposure and power guide explains the measurements needed before committing to a lighting arrangement.

The LS-LSC-0131 is a line-light candidate to assess for coverage and integration. Confirm its suitability through an exposure and contrast trial.

Record reference images from both views with representative acceptable material, defects, and backgrounds. Check edge regions as carefully as the center. A classification threshold that compensates for poor lighting in one area can hide real variation elsewhere.

Choose components without assuming compatibility

Use the camera finder to narrow candidates by required color information, line length, interface, and mounting arrangement. A monochrome camera is not a substitute when the decision depends on hue, although it may suit silhouette or intensity-based inspection.

The MV-CL042-91GC-V2 is a color GigE example. The MV-CL081-41CM is a monochrome Camera Link example. They illustrate different acquisition paths, not an interchangeable pair or a confirmed solution for the worked throughput requirement.

The MV-GC1102IOL is a Camera Link frame-grabber candidate. Confirm camera-mode support, timing inputs, software support, and host requirements before selecting it for a system.

Use the accessories finder for acquisition hardware and the lighting finder for illumination candidates. Request written confirmation of the proposed operating configuration rather than inferring compatibility from separate product descriptions.

Before ordering, document the delivered pixel format, sustained line rate, exposure range, trigger behavior, timestamp meaning, and supported simultaneous-camera arrangement. Include the lens, cables, lighting controls, and output timing hardware in that review.

Validate the complete detection-to-rejection chain

Commission in stages so failures remain traceable. Start with isolated objects, progress to paired objects, and then increase stream density. Include acceleration changes, lateral drift, touching objects, and deliberately ambiguous associations.

Use this release checklist:

• Verify visibility with defects at several surface positions. Record regions that remain uninspectable rather than reporting complete coverage.
• Measure crossing-speed variation and compare reconstructed object dimensions against physical measurements.
• Check both-view timing with a shared observable event. Repeat after changes to acquisition settings or software.
• Measure arrival-time prediction error at the rejection plane, including the tails of the error distribution.
• Measure command-to-effective-rejection delay under representative pneumatic demand and pulse sequences.
• Log dropped lines, unmatched objects, ambiguous matches, queue age, late decisions, and commanded rejection windows.
• Measure missed rejects separately from acceptable objects removed by neighboring pulses.
• Repeat the trial at the intended sustained stream density, with normal processing and recording enabled.

The usual opposing-view solution is wrong when hidden defects must always be inspected, the stream cannot be separated sufficiently, or trajectory variation exceeds the available rejection window. Additional controlled views, singulation, a different presentation method, or a different sensing approach may be necessary.

For acquisition shortfalls, follow the camera throughput bottleneck guide before increasing buffer sizes. For component selection, start the camera configurator, or send the inspection width, smallest defect, measured speed range, and rejection distance through application support. Those measurements establish a defensible starting point for both imaging and timing.