Line tracking keeps handing you the same symptom: items pick cleanly in the middle of the belt, results turn inconsistent near the frame edges, or an item that clearly entered the view never gets picked at all. Two settings named limit search region and limit position region look like the same crop drawn twice, and mixing them up produces exactly those symptoms. This guide separates the two: what each one does to a frame, why one discards while the other rejects, and the margins that keep both correct. The numbers come from belt speed, frame period, field of view and item size, so you can set both regions from your own line's data. The systems topic page gathers the related tracking and region pitfalls.
What is the difference between a limit search region and a limit position region?
The limit search region crops the image before it reaches the vision algorithms. The limit position region checks each detected pick point afterwards to confirm the whole item sits inside the camera view; one bounds the search, the other validates the pick.
They act at different stages of the same frame. The search region is applied before detection: pixels outside it never reach the matching, blob or code-reading tools, so a candidate outside the band cannot be found at all. The position region is applied after detection: the tracker already holds a candidate, and the region decides whether that candidate may become a pick.
The ordering explains the two different symptoms. A search region drawn too tight hides items completely; a position region drawn too tight finds items and then drops them. Both read as inconsistent counts in your reports, but knowing which border is mis-set tells you which one to move.
What does the limit search region cut from the image?
It cuts the picture down to the belt band before any algorithm sees it. An item outside the crop is never detected at all, and every tracked frame carries less data to search.
Size the band from belt geometry and trigger timing, then check the pixel count. Take a 5 MP frame as 2448 × 2048 pixels, which is 5,013,504 pixels in total. Suppose the belt occupies the central 80% of the width and the middle 700 of the 2048 rows. The crop keeps 0.80 × 2448 ≈ 1958 columns and 700 rows. The algorithms then search 1958 × 700 = 1,370,600 pixels, which is 27% of the frame (1,370,600 / 5,013,504 = 0.273).
Search cost scales roughly with the pixels the algorithm must visit, so the crop trims that stage by about 73%. If a full-frame search takes 24 ms on your processor, the cropped band lands near 6.6 ms (24 × 0.273). That headroom is what lets a modest on-device processor hold the tracking rate. When acquisition keeps beating the vision stage, the fixes in acquisition faster than processing: fix the frame rate gap start with exactly this crop.
What does the limit position region check on each pick point?
It verifies, after detection, that a found pick point sits far enough inside the view that the whole item and every sub inspection area stay fully within the image. Candidates that fail are discarded.
A pick point alone is not enough to inspect anything. The item extends around the point, and sub inspection areas — a seal check at each end, a code read near one corner — need their full area inside the image to run at all. Let a sub area hang over the frame border and its inspection fails, so the item leaves as a failed result even though the part itself is sound. The position region prevents that outcome by dropping the candidate before inspection runs, provided the region was drawn around the item envelope and not just the pick point.
The check is only as good as the position underneath it. A pick point found by a poorly tuned search carries its own error into the region test. That is why pattern search sensitivity vs accuracy: tuning position detection belongs in the same setup pass.
Why is an item outside the position region discarded rather than rejected?
Because a discard removes the position from this frame's pick list without registering a rejection. The overlap filter then has nothing to cancel, and the same physical item can be picked normally on the next frame.
Rejection is a recorded outcome; discard is a silent drop. Line tracking carries positions from frame to frame, and the overlap filter compares each new detection against the positions it already holds, so two picks never land on the same spot. Record an out-of-position item as rejected and its position stays in that memory. When the same physical item reappears, fully inside the view, the filter cancels the new detection at that same position, and the item is never picked on any frame.
Discarding breaks the chain. Nothing is registered, so the next detection at that position proceeds untouched, and the item gets its pick once it is fully visible. The same carry-over of tracked positions across frames is what conveyor target acknowledgment at tracking position rollover covers at the rollover boundary.
These regions live on the vision side only. The robot keeps its own taught tracking frame and converts each pick point with conveyor data arriving between frames; how line tracking uses a taught frame and timely position on the robot side covers that half of the handoff.
How much of the image should the search region cover?
Cover the full belt band plus one frame of belt travel. At 300 mm/s and 50 ms frames that is a 15 mm margin past the outermost item edge you intend to accept.
The frame period follows your trigger setting, not the camera's maximum rate; industrial camera trigger modes compared: free run to PTP scheduled maps the modes from free run to PTP-scheduled capture. With the period known, every margin above comes from one multiplication:
| Belt speed | Frame period | Travel per frame |
|---|---|---|
| 150 mm/s | 50 ms | 150 × 0.050 = 7.5 mm |
| 300 mm/s | 50 ms | 300 × 0.050 = 15 mm |
| 600 mm/s | 50 ms | 600 × 0.050 = 30 mm |
Fall below that margin and a new item enters the frame undetected, able to cross the pick window before the tracker first sees it. Grow far above it and you pay search time on empty belt. On a depth sensor the same placement question moves to the point cloud, which 2D vs 3D camera ROI: where to set it and which one wins settles.
How far inside the view should the position region sit?
Inset the position region from each image edge by half the item's longest dimension plus your tracking tolerance. For a 60 mm item with 10 mm of tracking error that is 40 mm, leaving a 400 mm pick band across a 480 mm field of view.
The arithmetic is direct: inset = 60 / 2 + 10 = 40 mm, and usable band = 480 − 2 × 40 = 400 mm. Take the tracking tolerance from measurement — robot follow error plus calibration residuals read off your pick-error log — not from a rule of thumb; the 10 mm here stands in for that measured value.
| Tracking tolerance | Position inset | Usable band (480 mm FOV) |
|---|---|---|
| 5 mm | 60 / 2 + 5 = 35 mm | 480 − 2 × 35 = 410 mm |
| 10 mm | 60 / 2 + 10 = 40 mm | 480 − 2 × 40 = 400 mm |
| 20 mm | 60 / 2 + 20 = 50 mm | 480 − 2 × 50 = 380 mm |
An item whose pick point clears the inset but whose end sits a few millimetres into the discard zone is discarded, correctly: that is the geometry that would fail a sub inspection at the frame border. Discarding it costs one frame of belt time; inspecting it at the edge would cost a false reject.
Which hardware runs tight regions fast enough?
Run single-camera line tracking on a vision sensor or smart camera when the cropped band keeps the search small. Move to a vision controller with one LAN port per camera once the cell runs several cameras or heavier algorithms.
A tight band pays back most where the processor is modest, because a smart camera runs inspection on the device and every cropped pixel is CPU time returned to the tracking budget. The MV-SC3050XC-16M-ATN-H is a 5 MP color vision sensor with a global shutter, M12 mount and Fast Ethernet output, suited to a single-view belt station. The MV-SC6050C-08M-XBN is a 5 MP color smart camera with a global shutter, an 8 mm lens and a built-in multi-color light source. For belt work stay on global shutter: a rolling shutter smears an item that moves during exposure.
- Interface
- Fast Ethernet
- Resolution
- 5 MP
- Interface
- GigE
- Resolution
- 5 MP
Past one camera, the tracking logic and inspection tools move onto a vision controller. The MV-VC2140-128G40-1T runs an Intel N-series processor with 8 GB of memory, 3-4 LAN ports and four built-in light channels, a compact fit for a one- or two-camera cell. The MV-VC3301H-128G66 steps up to an Intel Core i3 with 8 GB of memory, 5+ LAN ports, four built-in light channels and the VM6200Pro license included for the vision software.
- LAN ports
- 3-4 LAN ports
- USB 3 ports
- 2 USB 3 ports
- LAN ports
- 5+ LAN ports
- USB 3 ports
- 3-4 USB 3 ports
How do you verify both regions on a running line?
Overlay both regions on a live image, run a full disk through at production belt speed, and read the results. Detections must stop at the search band border, and edge items must yield neither a pick nor a rejection record.
Four checks close the loop:
- Confirm the search band holds the belt plus the travel margin from the table above, and that no detection ever appears outside it.
- Confirm the position region sits inside the search band by at least the item inset, so every detection receives the position test even when it ends in a discard.
- Park one item half outside the position region: it must produce no pick and no rejection record. Move the same item fully inside: it must be picked on the following frame, which proves the discard left nothing for the overlap filter to cancel.
- After a camera swap or cable change, re-check which image source each region is bound to. Region errors that survive correct geometry usually trace to the source binding, as ROI image source binding: fix inspection region errors walks through.
If both regions behave and picks still drift, the fault sits outside them, in trigger timing, calibration or robot-side tracking. To size the camera and controller for your own band, start from the categories above or go straight to the system configurator with your belt speed, frame period and field of view.
Questions
Does shrinking the search region change tracking accuracy?
No. The crop only bounds where the tracker looks; it does not alter pixel resolution, calibration or matching precision. A too-tight band fails as missed detections instead: items appear late in the results or not at all because they were cut away before any algorithm ran. Widen the band until edge items appear in the results, then stop adding margin.
What happens if the position region is drawn equal to the search region?
You lose the edge picks. Items in the outer band are still detected, but they sit inside the discarded zone, so they never reach the pick list even though every sub inspection area is inside the view. Keep the position region inset by half the item plus tolerance and let the search band stay wider.
Do both regions need updating when the belt speed changes?
Only the search band does. Travel per frame is belt speed times frame period: 150 mm/s at 50 ms gives 7.5 mm, and 600 mm/s gives 30 mm, so the band margin must grow with speed. The position inset depends on item size and tracking tolerance, which a speed change does not move.
Why does the result count drop after the search region is narrowed?
An item outside the search region never reaches the vision algorithms. It cannot appear in the results as rejected, discarded or picked; it is never seen. Compare detection counts with the item count on the belt: a gap concentrated at the frame edges means the band is narrower than the item stream, so widen it until the gap closes.
Can the position region act as a quality gate for the item?
No. It checks only whether the pick point, the item envelope and any sub inspection areas fit inside the view; it carries no information about quality. Items inside the position region still pass or fail through the inspection tools that run afterwards, so keep the reject decision in a separate inspection step.