Guest Speaker: Dr. Sanjay Chawla takes on the use artificial intelligence for the development of smart cities.
With the goal of expanding their AI/ML knowledge base, participants engaged in lecture-focused sessions spanning multiple domains. During these sessions Sibaq Lahja participants enjoy weekly access to speakers who showcased real-world applications of artificial intelligence, and machine learning solutions.
During one of these sessions, Dr. Sanjay Chawla, Senior Research Director at Qatar Computing Research Institute (QCRI), shared what is currently being done in the field of AI as it pertains to the development of smart cities. Requirements for these applications can range from mapping, to traffic and accident detection, plus more!
Here a snapshot of what was covered during the presentation:
Extracting maps from GPS and Satellite Imagery
So, we all have experience using applications like Google Maps, and are all aware of how useful such applications are in our daily lives. However, if you were to consider how these applications are built, it boils downs to a lot of manual work. In other words, big corporations behind these apps have massive teams that go around in a car and capture all of the relevant information required to develop the maps that we use. But even with the most precise mapping solutions, roads, structures, buildings, etc tend to change over time, and it becomes a challenge to maintain accuracy as the information might not be as up to date in these applications as practically required.
This turned out to be a major challenge in Qatar due to the rapid development, and changes in the infrastructure and road network happening at the time. Hence, the idea to see what application of AI might be used to help improve the accuracy of maps generated using GPS and satellites. Integrating AI into this process could simplify the process and greatly improve map accuracy, thus keeping the information continuously up to date.
Hence, a custom-built deep learning model was developed by the institute to solve this challenge. In this case, the very first step was acquiring the necessary data. Satellite images were gathered from various sources along with information on road networks that was retrieved from OpenStreetMap, the place for everything related to maps. This data helped train their deep learning model which uses satellite imagery as inputs, and generates road networks as outputs. Not only this but they also developed an application that lets you extract maps using satellite imagery from anywhere in the world, all done in real-time!
Build high-precision traffic crash maps
Traffic crash maps are, as the name suggests, maps that tell you the probability of a traffic accident taking place in a certain area at any given time. Such maps tend to account for contributing factors such as structure of the road network, amount of traffic, and steepness of curbs that may or may not result in crashes/accidents.
Having an AI-powered solution to build high-precision traffic crash maps can be useful in multiple different ways. For example, it can be used in navigation applications where users can not only select the fastest routes but also the safest routes to their destination. Another great use would be for insurance companies for generating more precision and data-based insurance. Additionally, AI models can help promote a more interactive city planning process where predicting the probability of crashes if roads were structured differently. In other words, city planners can rely on these AI tools to help design safer cities, and understand where and how accidents can be reduced, or even prevented.
The input data for development of this kind of AI-powered solution included: road maps, aerial imagery, GPS traces and history of accidents in the region. And, in this example, the output data was a high-resolution crash risk map.
Optimizing traffic signal control with AI
Imagine yourself waiting for a traffic signal, for a long-long time, even though there is no traffic in either direction–this is a common occurrence when timed signals are deployed. Now imagine, instead of timed signals, you deploy AI. These more “intelligent” applications can greatly improve the ability to recognize the amount of traffic, and adapt accordingly.
Challenges like this also demand AI-solutions that can achieve optimal results. Reinforcement Learning is a type of machine learning best suited for situations like these. Concerned with how intelligent agents should take actions in any given situation Reinforcement Learning systems are applied with the goal of maximizing their cumulative reward. So, from observing data, gathering a simulation of real traffic data, and feeding this formation into a neural network model, an output can be generated that increases efficiency. To be specific, the output here is the action of the traffic signals turning red, yellow or green according to the optimization of the flow and state of traffic.
The QCRI teams often deal with real-world challenges. By collaborating with local stakeholders, the institute can focus not only on solving the problem, but also conducting original research for introducing better solutions into the process.
As there is considerable growth and development being seen in such areas, it is a great time to be working in data science and the AI/ML ecosystem. In essence, AI technology provides a powerful set of tools to “smartify” applications. But we also need to go beyond just predicting, into “decision making”. Applications and solutions can be used to not only predict things, but to also help guide decision makers through a more data-driven analysis.
However, making AI tools “interpretable” is still a work in progress. Many of these models are huge, with millions of parameters, which can make it difficult to figure out why any particular solution works. Similarly, if something does go wrong, it can be challenging to debug as well. But for now, with so much data being collected in various forms, there is a lot of room for development.
To learn more about Sibaj Lahja, click here for more information.
If you are looking for tools to support your AI/ML model development journey, become an aiXplain member today.
Click here to join!