SLAM Technology: The Silent Engine Behind Real Estate Media

SLAM technology, short for simultaneous localisation and mapping, lets a device work out where it is inside a place it has never seen while it draws the map of that place, with each estimate used to correct the other. In real estate capture it is the silent engine that lets scans taken room by room fit together into one model of the house, which a floor plan can then be drawn from.

The person carrying the device still decides how well that works, because every part of the method depends on what the device is shown and how it is moved.

Localisation Without GPS

Localisation is the device's estimate of its own position and of the direction it faces. Outdoors, satellites can supply much of that. Indoors they can't keep the error in check, and a survey of the field published in IEEE Transactions on Robotics in 2016 dates the rise of SLAM to robots moving inside buildings, where GPS could no longer bound how far wrong they were.

Without satellites, a device can count its own movement, a method called dead reckoning. Early robots did it with wheel encoders, and the same survey notes that position worked out from wheel movement alone drifts so quickly that it becomes unusable after a few metres. Modern devices combine sensors. A phone running Google's augmented reality platform tracks visually distinct points in the camera image and combines them with readings from its inertial measurement unit to estimate where it is and which way it faces.

Even then the estimate drifts, though far less: the 2016 survey puts the drift of recent visual and inertial methods below 0.5% of the distance travelled. What reduces drift, and can correct it, is observing the surroundings as well as the movement. The survey counts SLAM and its close relatives among the techniques in use from self-driving cars to phones, and treats vision-based SLAM on slow-moving household robots as a mature field.

Mapping From Ranges and Features

The map is the device's model of its surroundings, and what it is built from depends on the sensor. LiDAR measures distance with pulses of laser light. Each range, combined with the position and orientation of the instrument when it was taken, becomes a point with three coordinates, and the points together form what NOAA describes as a point cloud. A camera records appearance instead, and its map is made of features it can pick out and find again.

LiDAR SLAM suits buildings well. Researchers at Google described portable LiDAR combined with SLAM as an efficient way to acquire as-built floor plans in a paper presented at a robotics conference in 2016, and they found that drawing the plan as the capture goes on helps the operator judge its quality and coverage before leaving the site.

Each sensor has its own blind spots. Two rooms of the same shape can look identical to a laser scanner, while a camera may tell them apart by their appearance. Researchers call it perceptual aliasing when similar scenes in different places make a device think it is somewhere it has already been. Distinct shapes and features are what let it recognise a place it has seen before, and recognising places is the reason both halves of the name have to run together.

Why Both Happen at Once

To put a new measurement on the map, the device needs to know where it stood. Working that out, in turn, takes the map. SLAM resolves this by estimating the map and the device's position together, revising both as each new measurement comes in.

The step that holds the estimate together is called loop closure. When the device comes back to a place it recognises, it can reset the error it has built up since it was last there. Without loop closures, the survey says, SLAM reduces to odometry. Its figure of a building mapped from odometry alone looks like one long corridor, with two points that are close together in reality drawn far apart, while the map built with loop closures takes the building's actual layout.

Inside a SLAM system, a front end turns raw sensor data into features and matches them between views. A back end, built on geometry and optimisation, estimates the map and the device's path from those matches, and it can send a possible loop closure back to the front end to be checked before it is accepted.

Where SLAM Loses Its Place

The method assumes the world holds still while it is being mapped. According to the 2016 survey, that assumption holds in a single mapping run of a small space as long as people and objects aren't moving around, and current algorithms can be made to fail easily by fast motion or by surroundings that keep changing. Spaces with few distinct features, or with several areas that look alike, leave the device little to recognise. Full-height mirrors and glass are hard for a LiDAR sensor to measure.

This leaves the photographer in the driver's seat, since a missed position, a door closed halfway through or a camera swung between rooms weakens the data that every later output is made from. Before and during a capture, these are the conditions to look out for:

• doors opened or closed, or furniture moved, after the capture has started
• people moving through the rooms being captured
• sudden or fast movement of the device
• mirrors, glass and other reflective surfaces
• large open spaces, and very small or empty rooms
• long gaps between one position and the next, with no line of sight between them

The Map You Check Before the Next Scan

The Giraffe360 camera records HDR images, 360 images and LiDAR measurement data in one device, with a LiDAR scan range of 0.2 to 40 metres and about 1.5 million points a scan. The photos, floor plans, virtual tour and videos are all made from the same visit, so each scan's place in the house matters to every one of them.

On the camera, Scan View is where that place can be seen. You press Scan, then Start Scan, and step out of the room until the scan is complete. The camera calculates the scan's position from tracking and adds it to a real-time map, which shows how far the capture has got and which areas still need covering. Where a scan sits in the wrong spot, Realign Scan lets you move it with gesture controls, using walls and furniture as the reference, and the Help Centre's ideal is within 0.3 metres and 25 degrees. Change location sends a scan to another floor, and Floor View shows which floor the camera is on. Giraffe360 says well-aligned scans make processing more accurate and bring the project back sooner.

Giraffe360's capture guidelines ask for what the method needs: each scan visible from the previous one, short and direct moves between them (closer in Double Scan mode than in Single), alignment checked after every scan and fixed at once, and the floor changed on the camera before you carry it to a new level. The floor plan manual adds the landmarks: capture the shapes that make each level unlike any other, such as corners, nooks and staircases, because they help arrange what it calls the room puzzle. Setting the camera where it can see its last position, especially between rooms, gives a smoother tour and better room alignment on the floor plan.

The placed scans then carry through to what buyers open. In the virtual tour editor, Settings has a Mini-map section with Show mini-map and Show room area, for the navigation visitors see as they move through the virtual tour. In the 3D floor plan editor, setting Virtual tour integration to Regular puts tour pins on the plan, so a visitor can go from the model of the house into the tour.