May 19, 2016 at 10:01

Exploration and engineering: 3D-models of buildings, interchanges and quarries according to the photo



    At some point, a man in civilian clothes appeared on the doorstep with a whole hard disk of photographs of one engineering object (let's call it for definiteness an overpass over the railway). The task was reduced to the traditional “once built a long time ago, and then lost the original documentation”, and you need to restore the project with an error of a centimeter or two.

    To understand how good accuracy can be obtained in architecture and engineering surveys of this kind, we decided to shoot our 11-story office and create a 3D model. At the end of the work, we got a model of the building and measured a couple of windows on it. Then they went and measured the same windows in the real world - the average discrepancy was within a centimeter, a maximum of two.

    By the way, one of our partners had a similar task in some ways - it was necessary to remove a quarry with an area of ​​470 hectares and:
    1. Determine the volume of rock production.
    2. Calculate the volume of stock products.
    3. Get a surface to clarify the level of the hydraulic dump.
    4. Get a 3D model of the entire territory.



    Some of the routes were

    decided with the help of a drone, long nails, a GPS receiver and two packs of plastic plates for 5 rubles apiece.

    Tasks and vendors


    The task of obtaining a 3D model from the prototype of the object or the object itself is quite easily solved in the gaming industry and other similar areas by 3D scanning of the object. We have our own 3D scanners, but putting something larger than a dog into them seems rather complicated. Most buildings, alas, are larger than dogs. Of course, there are also handheld scanners and industrial technologies where you can drive even a helicopter, but this is still a fairly separate specialized area. Incidentally, it is also suitable for buildings. But the use of photogrammetry and drones is more advantageous in time.

    In the case of the building, the model is used the same as for the scanner, but there are nuances:
    • Both technologies assume the presence of many images from various points, and for each image there is meta-information with the location of the camera.
    • The scanner works in "sterile" conditions, so filtering and image preparation are not needed. For large shoots, lighting can be different (elementary - during shooting the cloud covered the sun), so algorithms require more explicit reference points.
    • The shooting of engineering objects is usually done from drones, helicopters or different heights, so the set of shooting points is much more limited, and plus we need to filter the images taken during the anti-ballistic maneuver of the platform jitter.

    The principle of operation is this: based on the metadata of photographs and the search for similar fragments of images, clearly defined reference points are built. As a rule, these are different angles, contrasting elements of drawings, borders of objects. Each pixel in the photo is searched for a color match in other photos based on the hypothesis of exactly where the desired pixel will be, taking into account rotation and changing the location of the camera.



    Found matches become key points. If there are at least three reference photos, this point is already built in the 3D model. The spatial coordinates are triangulated: directrixes are drawn from each survey point to the selected point, and their intersection gives the desired value.


    These planes are the design plans for shooting, and in the middle - a crocodile. Below is an example of how it looks for a building.

    Filtering techniques are also used to improve the signal-to-noise ratio. The most popular Levenberg-Marquardt algorithm (or ligament method) to refine the coordinates of points. Then, based on the control points, the remaining points are restored with some accuracy, and all this is covered by polygons.



    The next step is texturing. The 3D model is deployed in a plane, and then the spatial position of the point is aligned with the original photo to set the color.





    Our task was most suited to ready-made packages Photoscan from Agisoft, Pix4Dmapper from Pix4D, ContexCapture from Bentley.

    Photoscan is a Russian "semi-professional" package that is quite affordable, but difficult to use on farms. Pix4Dmapper is a Swiss product for processing aerial reconnaissance data by drones. It is considered one of the best solutions for working with orthophotomaps, point clouds and digital terrain models. As a rule, it takes metadata from drones well out of the box. Slow, dear. ContexCapture is a versatile French product. It is difficult to configure, but very convenient for work, because it knows how to work with very noisy data and lays well on large infrastructures, clearly industrial. In the end, we settled on it.


    The resulting holographic 3D model based on 290 photos. We used it for demonstration on 3D tables.

    Practice


    First of all, we started not with the building, but with tests in the laboratory. It was necessary to understand how many photos on a complex object are minimal and sufficient. It turned out that a very small number of photos, of course, leads to a deterioration of the model (this is understandable intuitively), but also a large number too. It is important not to overdo it, because with a significantly excessive number of source images, filtering algorithms begin to weird and “blur” the coordinates. This results in uneven contours of objects. Surely this is solved by precise tuning of the input parameters, but it is much simpler to select only suitable images of good quality in the right amount.


    Test crocodile

    On a cluster of 6 cars a complex terrain miscalculated overnight.

    As a result, all the work on the building took less than one working day. Naturally, the resulting model cannot yet replace the worked-out model created by the engineer manually, since there are quite a few errors or inaccuracies in the image in the layout. From here the third task grew - to do it beautifully, that is, to modify it by hand after miscalculation. The fact is that we showed the results to different people, and they reacted differently.
    1. We show the reconstruction of the archaeological find to the museum representative. The reaction is just wow.
    2. We show the excavation pit to the engineer - he is keenly interested in whether it will be possible to fly the drone from inside the building (it is impossible, laser scanners can be used correctly for this).
    3. We show the building to the architect - in awe.
    4. We show it to the founding company’s founding director: “What cartoonish sucks?”

    In general, the result at the exit from the calculation module is not the most aesthetic, although quite good. Let's just say that for a commercial result like architecture presentations, you definitely need tailored textures with clear colors and well-defined contours. For our measurement task, the first model was enough for us.

    Career


    While we were experimenting with the building, we found out that a similar task was already solved by our partners - Photometer LLC. Their task was to quickly and cheaply remove a quarry of 4 square kilometers:
    1. Determine the volume of rock production.
    2. Calculate the volume of stock products.
    3. Get a surface to clarify the level of the hydraulic dump.
    4. Get a 3D model of the entire territory.

    Of course, the task of calculating the volume of the warehouse promised excellent civil and not very application. To solve the problem, the DJI Fantom 3 Quadrocopter, 5 batteries, a laptop and a GPS receiver were taken for coordination on the ground. They decided to place “beacons” in place for greater accuracy: on 470 hectares, they took 24 white plastic disposable plates, which were fastened with hundreds of nails. The coordinates of each plate were taken from the receiver with the minimum possible error.


    Flight Plan A

    strong wind blew (about 10 m / s) and it snowed. It was difficult to fly, so it took about 5 hours to shoot, all 5 batteries were consumed, 27 GB of material was shot. The total length of the route is 45 km.


    Example image from footage

    In the same Context Capture uploaded 27 GB of video and coordinates in the WGS 84 system. We launched the video processing and the aerial triangulation process. Calculation - 2.5 hours on a dual-core office computer. Then another 28 hours for a 3D model. The machine eventually processed 1860 photographs and built a three-dimensional model of the entire quarry (470 ha). The necessary formats were obtained from this model: Las (point cloud), OBJ (textured model), FBX (three-dimensional graphics) and orthophotomap. At this stage, the processing of photographic material is over.

    1.5 GB Las was handed over to the customer to compare the results of aerial photography and the classical method used in the quarry. Unfortunately, they couldn’t open this file in place, because AutoCAD Civil 3D refused to show any signs of life when trying to “chew” it. It was decided to discharge the point cloud 300 times.



    The result - here in this simplified file, the difference between the work of a surveyor with a total station and aerial photography amounted to 0.5% with a mining tolerance of up to 5% (according to the norm, depending on the volume).



    The accuracy of aerial photography is higher, since the model is built on a surface with millions of points, while the surveyor picks them up significantly less. On the other hand, 3 days for all the work (or 1-2 days when calculating in the cluster) instead of a month, and from the equipment - an available drone for 100 thousand rubles, and it started almost from the knee.

    Summary


    Photogrammetry used to be used in geology, topography, and archaeological excavations. Then - after one epic shoot of the denouement to restore the project and a few more shootings to monitor the state of the project - it became famous in construction. Recently, photometry began to be used in the film industry. As you can see, something has changed since 1852, since the French engineer E. Lossed shot Paris from a balloon.

    References


    Tags:

    Also popular now: