AI for smart cities
Artificial Intelligence (AI) and Machine Learning (ML) tools and techniques are already being applied to smart city and neighborhood projects around the world. Specific application areas are only limited by the imagination of the public and city planners, such as,
- traffic management,
- road safety,
- water resource management,
- e-mobility,
- public safety and law enforcement
- newer automation system
- digital delivery of public goods and service
Artificial intelligence in smart cities – Business Going Digital
This is mission-critical because of the explosive scale of the data and the dimensionality of the problems that are inevitable with rapid urbanization. According to the United Nations Department of Economic and Social Affairs, currently, 55% of the world’s population resides in urban areas. This is expected to rise to 68% by 2050.
There are active efforts to lay down a digital backbone for managing such a massive confluence of human beings and their aspirations. The recent push in IoT focuses on putting billions of sensors (of various kinds and capabilities) around major cities of the world. This is the first step towards getting the data. But ingesting, analyzing, understanding, and acting on that ‘Big Data’ requires transformational changes in the digital and computing systems that need to be in place.
AI in computer vision for cities – challenges and ideas
Computer vision and intelligent scene perception are two of the most prominent applications as these fields have completely been taken over by modern AI algorithms and systems such as convolutional neural networks.
However, there are some tricky limitations and challenges to building high-performance and cost-effective AI systems with the kind of data that is generally available in a smart city project. At the most fundamental level, it boils down to the fact that traditional supervised learning models and approaches may run into data sourcing problems when they are applied to smart city application areas.
And, I am not talking about receiving and ingesting the raw data – images, and videos in this case. Thousands of cameras and image sensors are routinely being added to every corner of modern cities and public infrastructure. Of course, there are challenges related to the ingestion, storage, curation, and processing of such a massive amount of digital data. But even if an organization can manage all of that with relative ease, they have to tackle the problem of high-quality data labeling.
Human beings – commuters, school students, joggers, cyclists, pedestrians – everybody can become potential sources of image/video data through their smartphone.
It is easy to get them to source the raw data autonomously. But it is very hard to get them to label the data accurately, diligently, and consistently.
In this regard, there are already some novel ideas and initiatives from tech startups. It’s a great start.
In parallel, we can think of some ideas on how to tweak the fundamental AI/ML problem a little bit so that the massive influx of data can be put to good use in the near term.
Change the metric and use Data Meshing
To start utilizing the existing data and algorithms for immediate benefit towards smart city systems and projects, we can expand our definition of the metrics that are used to ‘teach’ the AI systems. And, for that, we have to go beyond a linearized view of AI model training and move between the planes (like in a three-dimensional mesh).
It is easy to get human beings to source the raw data autonomously. But it is very hard to get them to label the data accurately, diligently, and consistently.
Use the routine tasks AI is already great at
AI algorithms are already great at many routine jobs when it comes to computer vision,
- object classification (Is it a car, is it a trash can, is that a dog?)
- object identification and placement (where is it in the scene?)
- semantic and instance segmentation (grouping and tagging of the family of objects)
- semantic understanding of faces and human poses (age, gender, gait, movement style of human beings w/o being personally identified)
- semantic understanding of situations (is that a fire? an accident? a roadblock?)
We can use all of these…
But they are just the beginning of the stack
Often, AI-based startups and even big tech organizations spend a lot of energy on building the best performing Deep Learning (DL) models to obtain great accuracy on the abovementioned tasks.
Thinking beyond the box – can we take what we already have from publicly available pre-trained models (for example released by Uber or Google) and use them as one of the components of a larger predictive model?
To start with, we tweak the way these AI modules are used,
- we assume that their accuracy and performance is already good in an average sense – for a city or neighborhood that is not too different from the training set of the model, the output can be trusted
- we design them such that they can output hard numbers or statistics, instead of the probabilistic outputs for a standard DL model
Here is the idea. Note the addition of the new extraction layer at the output of the DL model and some instances of questions that we are asking through this layer,
- how many cars were there in that image? how many pedestrians?
- what was the distribution of ages and genders in that public gathering?
- what was the traffic density like at that time of the day?
Connecting to the larger mesh
Once we start getting these numbers and answering these questions, we have to enmesh them with other dynamic datasets that we can source from the smart city digital infrastructure or digitized records of a neighborhood.
- road and urban planning maps
- housing records
- crime and safety records
- weather data
- public information pertaining to schools, residential units, and businesses
These all go to the data mesh.
Asking the right questions
This is just a 20,000 feet level overview. The key to getting meaningful results out of this plan is to,
- scope the project
- ask the right questions
That means, setting the right mechanics and metrics for this predictive model. Some examples are,
- Is there a correlation between weather patterns, traffic density, and public safety?
- How does the distribution of vehicular distance and positions impact the accident rate, for a given weather condition and time of the day?
So, we will be using the approach of linking statistical distribution to ground-truth a lot.
Note, we are asking questions in terms of impact on quantitive variables like accident rate and public safety records that are supposed to be both,
- critical for the health and development of a modern city
- expected to be digitized and well-maintained by any smart city program
So, for our predictive model training, we depend less on the unpredictable citizen data labeling and more on the combination of the statistical output of pre-trained AI models and digital city infrastructure data.
We won’t get > 90% accuracy, but that does not matter
We have to be ready to accept the fact that we won’t get a strong predictive trend or signal-to-noise ratio all the time. The model accuracy (or any other metric for that matter) will not be > 90% like the great AI models (ImageNet) advertise to deliver.
It should not matter much.
In most cases of smart city projects, there is no AI-enabled model. No machine learning metric. Not a great many baselines to compare to.
No human being can do this kind of data meshing in her head and come up with a predictive model. There is no "human-level performance" to match up to. This is not a problem of identifying cats vs. dogs.
This is just an idea to push forward, embrace the gifts of modern AI models and large-scale data analytics software tools, and deliver value to the dream of bringing true digital transformation for tomorrow’s smart city.
we depend less on the unpredictable citizen data labeling and more on the combination of the statistical output of pre-trained AI models and digital city infrastructure data
Summary
In this article, we touched upon the concept of using the output of pre-trained AI models in a statistical manner – feeding their counts and distributions to a larger predictive model.
The core idea is to spruce up this predictive model with deterministic, well-calibrated data from modern smart city digital infrastructure.
We argued that this idea is a parallel approach while the data collection and labeling from citizen developers (human beings sourcing image or perception data through smartphones) are being perfected and improved upon.
We also talked about the expected challenges and performance limitations for such a messy data model. However, any discovery of correlation or statistically significant trend should be able to deliver higher value to the management of a smart city initiative over current approaches that do not embrace the full power of large-scale AI and Data Science.
Loved the article? Become a Medium member to continue learning without limits. I’ll receive a portion of your membership fee if you use the following link, with no extra cost to you.
