Enriching Digital Maps with Artificial Intelligence

2020-02-01

Researchers from MIT and the Qatar Computing Research Institute have developed a new model that uses satellite images to label road features in digital maps, which could potentially improve GPS navigation. Providing drivers with more detailed information about routes can help them navigate unfamiliar locations. For example, lane counting can alert drivers to lane deviations or merges, information about parking spots can assist drivers in planning ahead, and mapping bike lanes can help cyclists navigate busy city streets. Creating detailed maps is an expensive and time-consuming process typically carried out by large companies like Google. They attach cameras to vehicles to capture videos and images of roads in an area, which, when combined with other data, can create accurate and up-to-date maps. However, this process is costly, resulting in some areas of the world being overlooked. A solution has been proposed that utilizes machine learning models on satellite images to extract and regularly update labeled road features. However, roads can be obscured by objects such as trees and buildings, making this a challenging task. In a paper presented at the Association for the Advancement of Artificial Intelligence conference, researchers from MIT and the Qatar Computing Research Institute described RoadTagger, which uses a combination of neural network architectures to automatically predict the number of lanes and road types behind obstacles. When tested on digital maps of 20 cities in the United States, the model achieved an accuracy of 77% in predicting lane counts and 93% in inferring road types. The researchers also plan to train RoadTagger to predict other features, such as parking spaces and bike lanes. RoadTagger relies on a combination of convolutional neural networks and graph neural networks. The graph neural network model represents the relationships between connected nodes in a graph and has become a popular tool for analyzing social networks and molecular dynamics, among other problems. The model is end-to-end, meaning it takes raw data as input and generates output without the need for manual intervention. The researchers state that this combination of architectures represents a more human-like intuition. For example, if a four-lane road is partially obstructed by trees, only two lanes may be visible, but people can infer the presence of additional lanes behind the trees through observation.

RECENT AI NEWS

RECENT AI TOOLS