Robustness D555 PoE
Hello Realsense-family,
I have a question regarding the robustness of the new RealSense D555 PoE.
Currently, I am using the RealSense D456 camera with an additional IR filter.
The camera is deployed in a very challenging environment. The scene contains repetitive patterns and skylight windows that introduce direct sunlight. In addition, the camera needs to capture a large indoor space of approximately 20 m in width and 20 m in height.
Unfortunately, I have found the RealSense D456 to be too weak for these conditions. It struggles with repetitive structures, shows saturation issues in the IR image due to sunlight, and provides insufficient detection of near objects in High Accuracy mode.
Therefore, my question: I have seen that the RealSense D555 PoE includes a new vision processor board. Can I expect a noticeable improvement regarding the issues described above? Would this camera be suitable for such an environment in terms of avoiding artifacts caused by repetitive patterns and saturation, and ensuring robust object detection?
Thank you very much for your support.
Best regards, Laurin
-
Hi Laurin Steiner D555 uses the same D450 camera module inside it that D456 has, so the depth / IR and RGB sensors are the same on both models.
In regard to the new-generation ASIC processor board in D555 - the Vision SoC V5 - the following description is provided for this new board: "Comprised of industry leading stereo disparity processing, motion estimation, vision DSP optimized for computer vision, and best-in-class Image Signal Processor (ISP) IPU7. The ISP IPU7 enhances the RGB with Geometric Distortion Correction (GDC) and Temporal Noise Reduction (TNR). Vision SoC V5 is best optimized for computer vision applications".
So it is certainly better processing technology on the D555. For your particular problems though, I would recommend the D455f. It is like a D456 without the water / dust resistant casing (which is not a problem if you are using the camera indoors) but with built-in IR filters that provide better results in locations with direct strong light and repetitive patterns.
https://www.realsenseai.com/products/d455f/
The IR filter used in the filtered RealSense camera models is the CLAREX NIR-75N. This filter is also available for purchase separately, enabling it to be fitted over the lenses on the front of a non-filtered RealSense camera. More information about the filter can be found here:
-
In regard to High Accuracy preset mode, this mode tends to over-strip detail from the depth image. Medium Density is usually a better choice as it provides a good balance between accuracy and the amount of detail on the depth image.
When the camera is facing strong light, defining a Region of Interest (ROI) on the lower half of the image can help.
Tilting the camera can help to break up repetitive patterns. Also, RealSense provide a PDF guide to mitigating repetitive patterns if you have not seen it already.
https://dev.realsenseai.com/docs/mitigate-repetitive-pattern-effect-stereo-depth-cameras
-
Hello MartyX,
Thank you for your response.
As I mentioned above, I am using my D456 camera with an additional IR filter. Unfortunately, this did not fully resolve the false‑positive issues caused by repetitive patterns and saturation.
I am therefore looking for a way to achieve more reliable obstacle detection in large‑area environments.
What is the common approach in such cases? Would adding a dedicated Vision Processing Unit be a suitable solution? Do you know of any other improvements I could try? I need a solution that will reliably work.
Thank you for your support.
Kind regards
Laurin -
You could try rotating the camera 90 degrees so it is oriented vertically on its side to see if the repetitive patterns and sunlight issues improve. If they do, you could then use the post-processing Rotation Filter to rotate the depth and color images 90 degrees in the opposite direction so the image is in its normal orientation whilst the physical camera is side-on.
It would also be worth increasing the Laser Power setting to its maximum value of '360' to see whether depth detection improves.
Whilst D456 has a wide horizontal field of view, it has a smaller vertical field of view, which would make it difficult to capture a 20 meter high room in the camera's field of view. A solution to that may be to add an additional D456 camera and mount it vertically above the first camera, positioning them close enough so that their two fields of view overlap. Doing so can also improve depth perception in the area where the FOVs overlap, as the combined IR dot projections from the two cameras creates a denser overall dot pattern that makes analysis of surfaces for depth information easier for the camera within the overlap area.
Placing a linear polarization filter (which is a different kind of filter from an IR filter) over the lenses on the outside of the camera can greatly negate glare from reflections. As any polarization filter can work as long as it is linear (and not the round type used in 3D glasses), they can be purchased inexpensively from stores such as Amazon by searching for linear polarizing filter sheet. The sheet could then be cut to size to fit the front of the camera.
-
Thank you for your answer. I’ve already considered some of your suggestions, while others are new to me.
Do you have an answer to my question:
Would adding a dedicated Vision Processing Unit be a suitable solution?Do you have any experience with Vision Processing Unit? Can it improve false depth, as false positives, saturation, repetitive patterns and on the other side robust detection of obstacles/objects.
-
The NCS2 (Neural Compute Stick 2) USB acceleration stick was able to be used with RealSense 400 Series cameras years ago to provide hardware acceleration for applications. It contained an Intel Movidius Myriad X VPU. As far as I am aware though, acceleration of RealSense applications was its main benefit and it did not provide enhanced depth analysis.
Yesterday RealSense announced a new D436 camera model that in combination with other technologies could provide advanced humanoid robot navigation and obstacle avoidance AI in real-time.
Please sign in to leave a comment.
Comments
6 comments