GeoMate: AI-Powered Mapping for Smart Mobility
How it Works: HD Mapping with Aerial Imagery
Autonomous vehicles and bots need HD maps to function, but for decades, the process of creating detailed and accurate enough maps has been a major bottleneck in the industry. Typical methods rely on lidar technology, which is incredibly expensive and time-consuming to collect and process – not to mention, the data capture emits a lot of carbon. Plus, the resulting maps are difficult to keep updated in an ever-changing urban environment.
One solution? Ditching the lidar altogether.
At GeoMate, we produce HD maps that are as precise and comprehensive as they need to be to support self-driving, without using lidar at all. Instead, our data input is aerial imagery, which is much more efficient and inexpensive to collect, allowing for rapid mapping and easy updates at just a fraction of the cost of a lidar-based HD map.
Here’s how it works…
Data Capture & Post-Processing
The first step is collecting the geospatial data inputs. We purchase high-resolution aerial imagery collected via plane or drone. Our data post-processing involves some light image processing.
In contrast, collecting lidar data involves sending fleets equipped with expensive, specialized sensors out into the field. These vehicles drive around, recording every detail within a given environment in the form of a point cloud. This data then undergoes extensive post-processing to turn each individual feature into a single point on a map. You start with millions or even billions of isolated points that make up an image, almost like pixels. Each object in the point cloud, comprised of many, many points, needs to be distilled down into a single location point. That’s a lot of computational cost!
Overall, this is of course highly time-consuming, expensive, and emission-heavy. And keep in mind that if you need to update your maps as the environment changes, you’ll have to do this whole process all over again. Both the collection and the processing are much more time-, cost-, and energy-efficient when working with aerial imagery.
Think of it like this: aerial mapping is comparable to adding a sweater to your online cart and clicking ‘buy’; lidar mapping is like driving out to a local farm, shearing a sheep, processing the wool, and then knitting a sweater by hand.
Feature Extraction
Our cutting-edge computer vision and machine learning technology then extracts features from the built environment. Our algorithms are trained to recognize and identify everything from road edges to stop signs to pavement markings. The AI categorizes each feature accordingly and identifies the location and geometry of the feature. The resulting feature map is accurate within 3-10 cm – the typical requirements for an HD map supporting autonomous driving.
We conduct a thorough round of manual quality assurance to ensure the accuracy of this identification.
The feature extraction process is similar when working with pre-processed lidar data. Once each feature is distilled into a single point,it’s common for an AI algorithm to identify and provide classification. The resulting HD map is similar in accuracy.
Semantic Attribution
In the next stage, GeoMate employs cutting-edge semantic attribution algorithms, powered by artificial intelligence, to assign meaningful labels and attributes to the extracted features. By leveraging deep learning techniques, the system gains the ability to recognize and categorize different objects with exceptional accuracy. The AI algorithms attribute semantic information to the extracted features, such as lane direction, speed limit, and other classifications. This semantic understanding enhances the richness and usability of the resulting mapping data.
Another round of manual quality assurance is conducted.
The semantic attribution process is also similar when using pre-processed lidar data.
HD Maps: The Final Product
Now we have an HD map!
So let’s recap: GeoMate’s big efficiency gains happen during data collection and processing. Using aerial imagery instead of lidar means less field work, specialized equipment, and drastically more efficient post-processing, ultimately leading to HD maps of the same caliber produced more quickly at a fraction of the cost. Plus, it’s quick and easy to update them.
Then, GeoMate’s specialized feature extraction AI allows us to quickly identify and classify all features of the built environment, which further accelerates the speed of the mapping process.
Overall, this represents a shift to unprecedented levels of efficiency and reliability in the HD mapping industry. For years, creating HD maps has been cost-prohibitive, delaying the journey towards driving autonomy. HD maps created with aerial imagery alone are the future of self-driving cars, robots, and other vehicles.
Want to learn more about GeoMate’s HD maps?
