3D Vision Cameras from China: How to Choose for Picking, Navigation and Inspection

Define the vision task first

A 3D camera is chosen for one measurement job. Write these down before you compare products:

• Working distance: the nearest and farthest distance from the camera to the objects, including the camera's mounting position on the robot or in the cell.
• Field of view: the area that must be covered at that distance, such as the full bin, the pallet footprint or the path ahead of a mobile robot.
• Detail: the smallest feature or position error that matters for the task, for example the gap between parts or the tolerance of a grasp.
• Surfaces: dark, shiny, transparent or textured parts, and whether they are mixed in the same scene.
• Motion and timing: whether the scene is still during capture, and the time available per capture in the robot's cycle.
• Environment: sunlight, factory lighting, dust, water, temperature and vibration.

These six points decide the depth principle, the camera family and how much the camera will cost to integrate. For how stereo, structured light and time-of-flight work, see structured light vs stereo vs ToF cameras.

Applications and the camera traits that matter

Read range figures correctly

Many datasheets give two ranges, and only one is useful for planning. Orbbec lists the Gemini 335 with a depth range of 0.10 to over 20 m and an optimal range of 0.26 to 3 m. The wider figure tells you where the camera can still return depth; the optimal range is where the maker recommends using it. Shortlist cameras whose optimal range covers your working distance with some margin.

Range can also depend on the camera mode. The Femto Bolt time-of-flight camera has a depth range of 0.25 to 5.46 m depending on the depth mode. Its wide field-of-view mode gives up to 1024 × 1024 at 15 fps, and its narrow mode gives 640 × 576 at 30 fps. Pick the mode your task needs, then read the range, resolution and frame rate for that mode only.

Figures from the Orbbec product pages for the Gemini 335, Gemini 336 and Femto Bolt, and the Gemini 330 series Ethernet datasheet for the Gemini 335Le.

The table also shows that one camera family can come in several builds. The Gemini 335, 336 and 335Le share a depth resolution but differ in optical filter, baseline and connection. Choose the build for the lighting and cabling on your robot, not only the family name.

Compare accuracy on the same terms

Accuracy is the figure most often compared wrongly. Makers express it as a percentage of distance, a fixed error in millimetres, or a combination, and they measure it with different methods. Orbbec, for example, states the Gemini 335's spatial precision as a root mean square error below 1.5% at 2 m. That describes depth noise at one distance under the maker's test conditions. It is not the error you will see when a robot picks a part.

To compare two cameras:

1. Convert each figure to your working distance, using the maker's own definition.
2. Check whether the figure describes precision (noise, repeatability) or trueness (offset from the real position).
3. Ask for the target, distance and lighting used for the measurement.
4. Add the error of hand-eye calibration and robot positioning, which usually matters as much as the camera.

Interfaces, software and mounting

A camera that cannot be integrated quickly is expensive whatever its price. Check:

• Connection and cable length: USB is simple on a desk; Ethernet with locking connectors is easier to route through a robot arm or a vehicle. Confirm the maximum cable length and whether power comes over the same cable.
• Host platform: operating systems and processor architectures the SDK supports, including ARM boards on mobile robots.
• SDK and drivers: languages, ROS or ROS 2 packages and their maintenance status, and sample code for point clouds and calibration.
• Synchronisation: hardware trigger and multi-camera synchronisation if you combine cameras or sync with a LiDAR.
• Picking or measurement software: whether the maker supplies application software, its licence model and the robot brands it supports.
• Mechanical and environmental: mounting holes, weight, IP rating, operating temperature and vibration rating for your installation.

Test with your own parts

Datasheets cannot tell you how a camera handles your parts. Before ordering, send sample parts or a recorded scene to two or three makers and ask for point clouds at your working distance under lighting close to your site. Include the hardest cases: dark rubber, polished metal, transparent packaging and parts touching each other. Agree the success measure in advance, such as pick success rate, measurement repeatability or obstacle detection distance. For a structured sample process, see how to evaluate robot component samples.

Checklist for a 3D camera inquiry

1. Working distance range and the field of view needed at that distance.
2. Smallest feature or error that matters, and the cycle time per capture.
3. Part surfaces, ambient light and environmental conditions.
4. Robot or vehicle mounting, cable route and host computer.
5. Interface, SDK, ROS support and synchronisation needs.
6. Application software required and its licence terms.
7. Sample test plan with success measures.
8. Quantity, delivery date, warranty and the calibration tools included.

Sources

• Orbbec Gemini 335 product page
• Orbbec Gemini 336 product page
• Orbbec Femto Bolt product page
• Orbbec Gemini 330 series datasheet for Ethernet devices