One of the main problems of autonomous vehicles is navigating in a GPS-denied environment. In this post, we focus on car positioning inside tunnels and present state-of-the-art accuracy in inertial navigation. The novel approach was developed by ALMA engineers combining deep learning, inertial sensors, and classical signal processing methods.
Background: Navigation without GPS
GPS (or more generally, GNSS) is available on most modern mobile devices and we use it daily to navigate. But what are our options when GPS is not available? Tunnels, buildings, urban canyons, parking lots, and similar environments suffer from zero (or limited) GPS service. Of course, other sensors can be used to address this problem; cameras for visual navigation, Wi-Fi (or specialized beacons) signal for various forms of triangulation solutions, and other similar techniques.
But why bother with those computationally demanding and hardware-based solutions when most of us hold an IMU sensor in our hands at every single moment? Don’t tell me you didn’t know!
A sensor that exists in almost any smartphone is the inertial measurement unit (IMU). It consists of an accelerometer and a gyroscope. The two components measure acceleration and angular velocity which are relatively noisy and obtains measurements relative to the smartphone reference frame (and not the vehicle or user). Hence, there is a lot of work to do to obtain a reliable positioning solution from a handheld IMU. But the cool thing is that it is always available and at a very high rate (in most cases it is more than 100 samples per second).
Tunnel Experiments with ALMA’s Positioning
To validate our solution, we ran tests in the Carmel Tunnel which crosses Haifa, Israel. The total length of the tunnel is 6.5 km, and it is made of two underground segments and a single open section where GPS is available. We focus on the first covered segment which is 3.2 km long and without GPS reception.
We used a Galaxy S21 device to collect IMU recordings at the maximum rate of 500 Hz. GPS position and heading were collected whenever a signal was available. The mobile device was firmly attached to the dashboard of the car at the drive start at an angle convenient for the driver’s use. ALMA’s algorithm was applied to the IMU and GPS (with only partial availability) measurements to obtain an accurate position of the car. After less than 20 seconds of driving the phone installation angles are detected and a fused inertial-GPS position is displayed to the driver completely replacing raw GPS data. Position, speed, and heading become available at a fast rate of 20 samples per second (GPS is only available at 1/20 of that rate) whether the car is in motion or not. Note that GPS speed and heading measurements are only available when the car is in motion. If the phone is removed from the dashboard fixture, the algorithm switches to regular GPS mode until the fixture is restored.
Over 10 validation drives were conducted across the Carmel tunnel in which the GPS signal disappears, and the algorithm transitions to fully inertial relying solely on IMU inputs.
Red color shows the raw GPS position, blue color shows ALMA’s IMU-based solution. Inside the tunnel where no GPS signal is available, ALMA comes to the rescue. GPS signal is restored at the tunnel exit and a difference (between ALMA and the restored GPS) of 18 meters is calculated. The underground section of the tunnel is 3.2 kilometers long.
What About WAZE’s Beacons?
Let’s compare the above results with WAZE’s solution that rely on beacons. On their website, WAZE state that they need to install approximately 26 beacons per kilometer inside the tunnel (a unit cost $30). Meaning that for a 3.2 km section, a total of 83 beacons must be installed, about $2500 total cost, before considering the maintenance cost (link to Waze website).
The alternative to positioning using GPS/cameras/beacons/LiDAR/etc. are the widely available IMU sensors. However, measurements from those noisy and low-cost sensors are not easily turned into an accurate positioning solution. Add to this the challenging environment of mobile devices and general reference frames and the task might seem impossible. A SOTA result was presented for inertial navigation inside one of Israel’s longest tunnels. Positioning accuracy at the tunnel exit was shown to be comparable to that of GPS and achieved without added hardware and with a light computational load.