In human sciences stereo vision or stereoscopic vision is the ability to obtain a spatial visual impression with both eyes through image comparison. When viewing an object, each eye looks up from a different angle. Each eye filters the information and sends it to the brain, where both visual impressions are processed into a combined image. This is what creates our three-dimensional depth perception.

Computer stereo vision mimics human depth perception. With a stereo camera two images are recorded synchronously and compared. The distance information between the camera and the object observed is lost by taking an image with a conventional camera, but this depth information can be recovered by several images taken from different known camera directions and their comparison.

Stereo matching is the applied process in computer vision for stereo image comparison. The operation is based on finding all the pixels in the stereo images that correspond to the same 3D point in the captured scene.

In order to obtain the 3-dimensional points, once all correspondences were found, a triangulation has to be calculated by taking into account the intrinsic and extrinsic geometry of the cameras and their calibration.

With Nerian’s SceneScan this process is performed on an FPGA, speeding up the calculation process by magnitudes and providing real-time 3D-measurement!

In the context of stereo vision and depth image evaluation, disparity is the offset of the positions thatan object occupies on two different image planes.

If we stick to the example of human vision, our eyes are located at different lateral positions, which causes two slightly different image perceptions. So with the left eye we see more of the left side of an object, while with our right eye we see more of the right side. This positional difference is called disparity. You can check that yourself if you, for instance, close one eye and look at an object. Then close your eye, while open the other one. You will see that the object is detected at a slightly different position. The nearer an object, the higher the difference.

For determining the disparity in machine vision, stereo matching makes use of comparing all matching pixel columns.

The disparity range in stereo vision specifies the overlapping image region that is searched for pixel correspondences. The larger the disparity range is selected, the more accurate the measurement results will get. But a large disparity range causes a high computational load, and thus lowers the achievable frame rate.

The selection of a small disparity range is suited for high speed measurement applications, while the selection of a large disparity range is appropriate for closer range measurements and if higher accuracy is required. Nevertheless, a larger disparity range increases the invalid left border where the camera image cannot be compared.

With SceneScan we offer a configurable disparity range from 32 to 256 pixels. We recommend the following combinations of resolution, disparity range and frame rate:

For Scarlet we have the following combinations:

Disparity is the apparent displacement of pixels in a stereo image pair (see What is disparity?). The disparity map is computed from this image pair and contains the depth information of each recorded pixel. Structurally it is a 2D image, so many 2D image processing algorithms are applicable.
In contrast, a 3D point cloud provides 3D coordinates for each pixel, which allows actual 3D measurements. Nevertheless the 3D output requires more complex algorithms and more data than disparity maps, causing slower image processing.
For faster results, working with disparity maps should be preferred over 3D point clouds.

SceneScan and SceneScan Pro are embedded image processing systems for real-time stereo matching. SceneScan connects to a dedicated stereo camera or two industrial USB cameras, which are mounted at different viewing positions.

SceneScan sends out trigger signals in order to synchronously capture stereoscopic image pairs from the connected cameras. By correlating the image data from both cameras on the integrated FPGA, SceneScan can infer the depth of the observed scene. The computed depth map is transmitted through gigabit ethernet to a connected computer or another embedded system. SceneScan captures up to 100 frames per second, providing depth measuring in real time.

Our Scarlet 3D depth camera works on the same principle, but combines a 3D stereo camera and image processing in one device. A particularly powerful FPGA allows a processing performance of up to 120 fps, over 70 million 3D points per second and that at a resolution of up to 5 megapixels.

In contrast to conventional 3D sensing solutions, our sensors passively. This means that no light needs to be emitted for performing measurements. This makes our stereo vision solutions particularly robust towards the illumination conditions, and it facilitates long-range measurements, the use of multiple sensors with overlapping field of views, and a flexible reconfiguration of the system for different measurement ranges.
Furthermore we provide faster, higher-resolution and higher quality depth maps, because we can harness the high computational capabilities of a FPGA and use high-quality image sensors. Because all image processing is done on SceneScan or Scarlet, there is also no computational load on the host PC.

Our big advantage compared to LiDARs is that we have a much higher vertical resolution (typical LiDARs only measure in up to 64 rows). In some applications this can result in objects with a low height not being detected.

• Lower power consumption
• Smaller size
• Faster results and more reliable timing because no other operating system or software are competing for resources

SceneScan and Scarlet deliver the stereo matching results in the form of a disparity map from the perspective of the left camera. The disparity map associates each pixel in the left camera image with a corresponding pixel in the right camera image. Because both images were previously rectified to match an ideal stereo camera geometry, corresponding pixels should only differ in their horizontal coordinates. The disparity map thus only encodes a horizontal coordinate difference. Additionally the left camera image is also output. 3D point clouds are possible as well.

Two different models exist for the given image processing system: SceneScan and SceneScan Pro. Both models provide the same functionality, however, SceneScan Pro has significantly more computational power when compared to SceneScan. This means that SceneScan Pro can process a given input stereo image much faster than SceneScan.

Thanks to the additional processing power, SceneScan Pro is also capable of processing higher image resolutions, color images and larger disparity ranges. Due to these benefits, SceneScan Pro can achieve a higher measurement accuracy than SceneScan.

You can find a brief comparison between both 3D sensor devices on the SceneScan product page.

A detailed comparison of the achievable frame rates at different image resolutions and disparity range can be found here.

Yes, you can purchase SceneScan without our stereoscopic camera. Nevertheless you need a stereo camera configuration for usage. Thus you need to acquire third party cameras. For recommendations on cameras please see “Can I use third party cameras?”

The maximum frame rate that can be achieved depends on the image size and the configured disparity range. The following table provides a list of recommended configurations. This is only a subset of the available configuration space. Differing image resolutions and disparity ranges can be used to meet specific application requirements. Here is an overview for SceneScan:

Our recommendations for Scarlet are as follows: