D435i for agricultural use
Dear community,
I am an industrial engineer student.
For my master thesis I am looking for a 3D scanner that would be used for agricultural applications. A very accurate point cloud is required as it will be used for product analysis. I'm thinking of the D435i (or D450), although I'm not sure they're accurate enough to within a few millimeters. Can using multiple cameras increase accuracy?
When I think of agricultural applications I think of very dusty environments and a lot of vibrations and especially a lot of influence from sunlight.
When 3D Lidar laser scanners with mechanical parts do not seem robust enough, I have to look for an alternative. How well does the D435i handle in these harsh conditions?
Thanks in advance.
With best regards,
Nick
Electromechanical engineer student KU Leuven University of Bruges
-
Hi Nick Vanoost The 400 Series cameras have already been used successfully in outdoor agricultural applications, such as being mounted to agricultural vehicles.
If your camera will be travelling over rough ground that causes vibrations though, I recommend installing vibration damper pads on the mounting point of the camera. The depth sensing algorithm of the 400 Series is robust when experiencing vibration, but damping pads will help avoid disruption to the camera if the vibration is particularly severe.
Sunlight should not be a problem as the 400 Series cameras actually work even better in strong light, though you should avoid pointing the camera directly at the sun as this may saturate the infrared imager component and cause the data stream to freeze until the camera is reset.
The new D455 model has 2x the accuracy over distance of the D435 models. For example, the D455 has the same accuracy at 6 meters as the D435 models do at 3 meters.
The D455 also has internal hardware design refinements that make depth to color alignment easier. This includes both the depth and RGB imagers being mounted on the same stiffener and both depth and RGB having the same Field of View (FOV) size. The RGB imager of the D455 also has a fast Global Shutter for the first time on the 400 Series, helping to avoid blurring on the RGB image during motion.
Yes, having multiple cameras whose field of view overlap can help to improve the quality of depth data. In the areas of the observed scene where the fields of view overlap, there is redundancy in the depth data as the same area is being observed by more than one camera. The more cameras that you have, the less blind-spots that there will be.
The accuracy of the 400 Series cameras is less than 1% of the distance of the object from the camera. At 1 meter range from the camera, this would translate to an accuracy of 2.5 mm to 5 mm.
There are a range of methods by which accuracy may be able to be enhanced though, such as using a High Accuracy camera configuration "Visual Preset" that sifts the balance of the camera's configuration towards favoring accuracy over fill-rate, or Medium Density that balances accuracy and fill-rate. Information about Visual Presets (also known as json) can be found in the link below:
https://dev.intelrealsense.com/docs/d400-series-visual-presets
Intel also have a wide range of white-paper documents that can provide expert advice for improving results. The tuning depth cameras for best performance guide that is the first item in the list of papers in the link below is especially worth reading for a general overview of how improvements can be made.
https://dev.intelrealsense.com/docs/whitepapers
In regard to dust, there is the possibility of creating a custom 3D-printed protective housing for the camera. If you require an off-the-shelf protective solution to provide complete protection against dust and water, the RealSense partner company FRAMOS offer an industrial version of the D435 camera called the D435e that has IP66 protection. You should bear in mind though that as this is not an Intel product, FRAMOS and not Intel are responsible for technical support of the D435e.
-
Hello MartyX Grover,
Thank you for the previous feedback. The intel realsense D455 module will be very interesting for my thesis.
Lidar offers an accuracy of 25-60 mm. Which I think is too much for my application. the goal is to make a 3D scan of a straw bale (dimensions 2.5 x 1.2 x 90) with a very high accuracy. These measurements are made in the field on a moving tractor.
In the datasheet of the D455 I read that there is an absolute error of >2% up to 4 meters. According to the datasheet this will then be >80% mm at 4 meters. In your previous post, you wrote 2.5-5 mm at 1 meter. This is an acceptable accuracy for my application, but what did you rely on to get this value?I also hear a lot from stereo vision that processing this data requires an enormous amount of computing power and that, compared to lidar, much more data is produced from the module. What data amount exactly are we talking about with the D455 in mb / s maximum and average?
The compatibility with LABview is very interesting, the company where I work has a lot of experience with this software. How exactly does the connection with LABview work? Are there any limitations compared to ROS?
Thanks in advance.
With best regards,
Nick -
As far as I know there has never been a mathematical formula quoted alongside the 2.5 to 5mm accuracy values, so I do not have an officially confirmed calculation method to provide. These values are quoted officially as the depth accuracy range of the L515 at 1 meter in Intel's official online FAQ for the L515 though, so they are trustable.
Regarding LabVIEW, my understanding is that using the wrapper provides a similar range of capabilities to those available when using librealsense with C++ since the wrapper's VI-library wraps most of the core functions of the librealsense library file realsense2.dll.
https://github.com/IntelRealSense/librealsense/tree/master/wrappers/labview#introduction
In regard to computing power, 400 Series cameras contain a dedicated Vision Processor D4 hardware board to do computation that would otherwise have to be performed using a computer's hardware resources. This enables the camera to be used on low-specification, low-power computing devices. You can read more about the D4 in the PDF document linked to below.
Intel's white-paper document about multiple cameras has tables that show bandwidth use, so the column for '1 unit' should give some general estimates of what to expect from bandwidth usage for various stream configurations on a single camera. It has tables for connection to a hub and direct connection to a PC.
-
Hello,
I have come to the point that I am going to order a certain camera. I still have some doubts whether I choose the right model for my application.
The D415 is claimed to be 2x as accurate as the D435i. In addition, it is claimed that the D455 is 2x as accurate on range than the D415. Can I then assume that the D415 is similar in accuracy to the D455?
Ideally I want the best accuracy possible (D415) in combination with an IMU. In my application we want active stereo vision, so using the IMU the images are stitched together. In any case, the D455 seems suitable given the presence of the IMU. SLAM algorithms can be used here, but they probably have to be programmed yourself.
In the Intel range I also find the T265 tracking camera, which implements V-SLAM algorithms using an IMU. Is it a common setup for the D415 and a T265 to work together to create an active 3D scan? Or do I make it extra complex for myself and do I choose the D455 better?
Thanks in advance.
Regards,
Nick -
The D415 has around 2x less depth noise than the D435 models, which converts to around 2x better accuracy over distance when reading a chart of 'RMS Error' over distance. It also has an optimal depth accuracy resolution of 1280x720 compared to 848x480 on the D435 / D455 models.
The D455 has around 2x better accuracy over distance than the D435 models (not the D415). So the D455 has the same accuracy at 6 meters as the D435 has at 3 meters.
If the camera is being used on a moving agricultural vehicle though then I would advise against the D415 because it is not as suited to capturing fast motion as the D435 and D455 models.
If you need an IMU then this will exclude the D415 from consideration, as it does not have one. The D435 does not have one either, but the D435i does. So of the stereo depth camera models available, it would be a choice between D435i and D455.
You can get very good tracking data from a stereo camera IMU on its own. In robotics a T265 and a 400 Series would typically be paired together for the purposes of V-SLAM navigation with the T265 and obstacle detection and avoidance with the 400 Series.
If you are performing very close range scanning then I recommend a D435i, as the D435 models have a minimum distance of the camera from an observed object of 0.1 meters, whilst the D455 has a minimum distance of 0.4 m.
If the observed object is 0.4 m or more distant from the camera's position then a D455 will provide enhanced accuracy if you plan to do depth to color alignment.
Please sign in to leave a comment.
Comments
5 comments