The key terms of light field, 3D machine vision and vision-guided robotics – explained in brief, with links to matching products, solution packages and articles.
Marking objects in images, for example with boxes and class labels, so that a neural network can learn from them. The quality and consistency of the annotations largely decide how reliably the trained network detects parts later.
The distance between two viewpoints, for example between two cameras of a stereo or light-field system. A longer baseline produces larger disparities and therefore finer depth resolution at a distance; several baselines of different lengths, as in a light-field camera, combine precision with robustness against occlusion.
A robot takes randomly arranged parts out of a container, guided by a 3D camera. The software has to locate the parts, choose one that can be gripped without collision and plan the robot path – and the task gets harder as the bin empties and parts lie against the walls and in the corners.
A digital 3D description of a part’s geometry. In CAD-based matching, the software searches the point cloud for the shape from the CAD file and returns the pose of every match – without training images. Without a CAD model, a trained neural network can take over detection.
Measures a camera’s optical properties – focal length, image centre, lens distortion (intrinsic parameters) – and its position and orientation (extrinsic parameters). It turns pixels into metric measurements; in a multi-camera system such as LumiScan X, it also relates all cameras to each other.
The Linux-based automation operating system of Bosch Rexroth, running on ctrlX CORE controllers and industrial PCs. Functions are installed as apps; the LumiScan vision apps run on it directly and exchange data with the PLC via the ctrlX Data Layer.
A shared data space in which programs publish values and functions as named nodes and read those of others. In ctrlX OS it connects apps with the PLC; LumiScan Keryx provides the same data layer on Windows and Ubuntu, so the control code stays the same on all three platforms.
A square or rectangular 2D code that stores a lot of data in a small area and stays readable when partly damaged. It is the standard for direct part marking – lasered or dot-peened straight onto the component – and for traceability in production.
A robot unloads cartons, sacks or other packages from a pallet layer by layer. A 3D camera finds the reachable items, even in shifted or mixed loads, and the software decides the order in which to remove them.
An image in which each pixel stores a distance from the camera instead of a brightness. It is the direct output of most 3D cameras and can be converted into a point cloud with the camera calibration.
The shift of the same scene point between two images taken from different viewpoints, measured in pixels. Near points shift more than distant ones, so disparity encodes depth: Z = f · b / d, with focal length f, baseline b and disparity d.
The geometric relationship between two calibrated camera views: a point in one image can only appear along a known line, the epipolar line, in the other. This shrinks the search for corresponding points from the whole image to one line. Stacking the views of a light-field camera gives the epipolar plane image, in which each scene point forms a line whose slope gives its depth.
The area of the scene a camera sees at a given distance. It follows from sensor size, focal length and working distance, and together with the sensor resolution it determines the smallest detail the camera can resolve.
A camera standard of the EMVA: cameras describe their features, such as exposure time, gain or trigger, under uniform names. Software can then configure cameras from different manufacturers in the same way.
An interface standard for industrial cameras over Gigabit Ethernet. It allows long cables and standard network hardware; together with GenICam it makes cameras from different manufacturers interchangeable.
The position and orientation at which a gripper takes hold of a part. It is defined relative to the part, for example on its CAD model, and checked for reachability and collisions with the bin and neighbouring parts before the robot moves.
Determines where the camera sits relative to the robot: mounted on the robot arm (eye-in-hand) or fixed in the cell (eye-to-hand). Only with it can a position measured in the image be converted into coordinates the robot can move to.
Applying a trained neural network to new images in production. It runs on the production PC or controller, often accelerated by a GPU, and its speed feeds directly into the cycle time.
A laser projects a line onto the object and a camera views it at an angle; the line’s position in the image gives the height profile. A complete 3D image needs relative motion, for example a conveyor or a robot moving the sensor, one profile at a time.
The light field describes not only how bright the light at each point is, but also from which direction it arrives. A light-field camera samples it by recording the scene from many slightly offset viewpoints at the same moment; software derives the depth of every pixel from the differences between those views. LumiScan X records 13 views in a single shot.
A robot loads blanks into a CNC machine, press or test station and removes the finished parts. Vision locates the workpieces in drawers, trays or bins, so they need less precise pre-positioning.
Building pallets from cartons of different sizes and weights, as in order picking for retail and e-commerce. Software calculates a stable arrangement from the order data before the robot builds the pallet.
A computing model that learns patterns from examples instead of following hand-written rules. In machine vision, neural networks detect, classify or segment objects; trained on annotated images of your parts, they recognise them under changing conditions.
Finding objects in an image or point cloud and determining where they are – as a bounding box or as a full pose. Classical methods compare shapes or contours; AI-based detection uses neural networks trained on annotated images and copes better with varying appearance.
Reads printed, dot-peened or laser-marked text from an image and converts it into characters a computer can process, such as part numbers, batch or date codes. A confidence value per word lets the PLC reject uncertain reads.
An open file format for trained neural networks, so that a network trained in one framework can run in another tool. Models from the LumiScan Training Framework, for example, run in AI Object Matching and VGR Lite.
A vendor-independent standard for industrial communication: servers offer variables and methods, clients such as PLCs or SCADA systems read and call them, with built-in security. LumiScan apps can be reached over OPC UA through the data layer.
A rectangle around an object that rotates with it instead of staying aligned to the image axes. For elongated or tightly packed parts it describes position and orientation far better and includes less background.
Picking individual products from a conveyor, tray or work table and placing them at a defined target. With vision, the products need no precise pre-positioning; for simple, well-separated parts a 2D camera is often enough.
The industrial controller that runs a machine’s control program: it reads sensors, drives actuators and coordinates the cell. A vision system typically receives its trigger from the PLC and returns results such as positions, codes or pass/fail decisions.
A camera that records the light field. Designs range from a single sensor behind a microlens array to an array of individual cameras. The microlens design is compact but divides one sensor’s resolution among the views; a camera array keeps the full resolution per view and offers longer baselines.
A set of 3D points with X, Y and Z coordinates, often with a colour or intensity value, that describes the surfaces a 3D camera has captured. Robot guidance software uses point clouds to locate parts, check for collisions and plan grip points.
Every depth measurement contains noise and stray points, for example from reflections or at edges. Filters remove points that lie too far from their neighbours (statistical outlier removal, SOR) or have too few neighbours within a radius (radius outlier removal, ROR); smoothing evens out surfaces without blurring edges.
The position and orientation of an object in space, usually six values: three for translation (X, Y, Z) and three for rotation. Robot guidance software reports the pose of each detected part so the robot can approach it correctly.
Mirror-like reflection on polished, oily or painted surfaces: light leaves at the angle at which it arrived instead of scattering in all directions. A single camera then sees a highlight or nothing at all at that spot, which causes holes and false depth values in many 3D methods. Several viewpoints, as in a light-field camera, make it likely that some views see the point without a highlight.
Depth from two images of the same scene taken a known distance apart, as in human eyesight. It needs surface texture to match points between the two views; on uniform or reflective surfaces, active stereo adds a projected pattern. A light-field camera extends the principle to many views.
A 3D method in which a projector casts known patterns, such as stripes, onto the scene and a camera measures how the surface distorts them. It is very precise on matte surfaces but usually needs several images per measurement, and shiny surfaces can reflect the pattern away from the camera.
Finds a known object by comparing the image with a small reference picture of it, the template. It is fast, needs no training and suits parts whose appearance hardly changes.
A 3D camera emits modulated or pulsed light and measures, for each pixel, how long the light takes to return. ToF cameras are compact and fast but resolve depth more coarsely than triangulation methods, and reflections from shiny surfaces or multiple light paths distort the measurement.
Robot applications in which a camera determines where a part is, so the robot can handle parts that are not in a fixed position. The vision system passes the part’s pose or a grip point to the robot controller – in robot coordinates, thanks to hand-eye calibration.
Instead of fixing a workpiece precisely in a dedicated fixture, a 3D camera measures where it actually lies, and the taught robot path is transformed to that pose. This saves fixtures and speeds up changeovers, for example in welding or machining.
The distance between the camera – more precisely the front of its lens – and the object. It sets the field of view and, for 3D cameras, the measuring range in which depth is accurate; in bin picking it has to cover the bin from the top down to the floor.
The technical glossary of the LumiScan documentation explains many more terms on installation, configuration and interfaces.