Google has come up with a solution to the parking problem that has been complaining for years

Recently, Google added a new feature to Google Maps in an update for Android devices - predicting parking conditions at the destination.

Users only need to query directions in Google Maps to see a new icon. If Google predicts that your destination may be difficult to park, you will see a colored dot appear on the map with the letter P. The difficulty of parking will be divided into three levels: "limited parking spaces", "medium" and "easy".

If Google predicts that your destination may have difficulty parking, a colored dot will appear on the map with the letter P.

An official Google blog post said the feature is currently available in 25 of the largest urban areas in the U.S. But the company is looking to expand the feature to more cities and countries in the future, as well as to the Google Maps app on Apple's iPhone.

Waze, another map app on the Google platform, also provides similar services for many major cities in the United States, but Google Maps uses a slightly different approach. Waze uses parking data information integrated by INRIX when estimating parking spaces, but Google claims that its parking predictions are based on anonymous data obtained by users who actively share their location history.

On February 3, Google software engineers James Cook, Yechen Li and researcher Ravi Kumar jointly released a research report detailing the principles of this application.

To provide this parking prediction function, engineers need to solve many problems - parking conditions are complex and changing, and there is almost no real-time parking space information; even in some areas where there are Internet-connected parking meters, this data does not include information such as illegal parking, permitted parking and early departures; roads can only provide two-dimensional images, but the structure of parking lots themselves is more complex; the supply and demand of parking spaces change instantly, and even the best system may not be able to update in time.

To address these issues, Google's team combined crowdsourcing and machine learning to build a system that can provide information about parking difficulty and even help users decide which mode of travel to take. In pre-release experiments, they found that the number of clicks on the travel mode button increased significantly, indicating that after users have obtained parking difficulty information, they will be more inclined to consider public transportation instead of driving. [Editor's note: Crowdsourcing refers to the practice of a company or organization outsourcing work tasks that were previously performed by employees to a non-specific (and usually large) public network on a voluntary basis. ]

To design an algorithm to solve the parking problem, three things are needed — ground truth data obtained through crowdsourcing, a suitable machine learning model, and a strong feature set to train the model.

Ground truth data

In machine learning solutions, collecting high-quality real-time data has always been a key challenge. The approach taken by the Google team was to ask drivers whether they had encountered parking difficulties. However, they soon discovered that such subjective questions often resulted in conflicting answers: for the same location at the same time period, some people answered that it was "easy" to find a parking space, while others answered that it was "difficult." By switching to an objective question like "How long does it take to find a parking space?", the credibility of the answer was greatly improved, allowing crowdsourcing to generate a high-quality, real-time data set of more than 100,000 answers.

Model features

With the available data, the next step is to select the features used to train the model. This project uses anonymous aggregate information provided by users who like to share their location as a key source of information for evaluating real-time traffic conditions, high-traffic time periods, and visit duration.

Google researchers said in the report that they soon discovered that even with the required data, they would still encounter some unique challenges. For example, if someone parks in front of their own door or in a private parking space, the system should not mistakenly assume that the parking space here is available. Users arriving by taxi may create the illusion that there are many parking spaces in front of the door. Similarly, public transportation users may be thought by the system to be parking at the bus stop. These wrong perceptions will mislead the machine learning system.

Therefore, the system needs to be designed with more powerful aggregation features. One of the features was inspired by the Mountain View area where Google is located. If Google Navigation finds that a large number of users are driving around the city center during lunch time, it means that parking spaces are difficult to find.

If Google Navigation sees a lot of users driving in circles around the city center at lunchtime, that's a sign that parking is hard to find.

Google researchers are thinking about how to use these "clues" about parking difficulties as a feature for training. The researchers compared the time it takes users to reach the destination directly with the time it takes to actually reach the destination by circling, parking, or walking, and aggregated the difference between the two. If there is a significant difference in the time most users take between the two, it is considered that they have encountered a parking difficulty problem.

After that, the Google team continued to develop more features: specific destinations, parking locations for walks, time and date of parking (for example, users park very close to their destination in the morning and far away during rush hour, what should we do?), historical parking data, etc. In the end, they got about 20 different features. After that, it was time to adjust the performance of the model.

Model selection and training

For the above features, the researchers used a standard regression machine learning model. There are several reasons for this choice: first, the principle of logistic regression is well known, and it is resilient to noise in the training data; second, the output of these models can be interpreted as the probability of difficult parking, which can then be mapped into descriptive terms such as "limited parking spaces" or "easy parking"; third, it is easy to understand the impact of each specific feature, which makes it easier to verify whether the model is reasonable. For example, when the researchers started training, many people thought that the above "clues" function would be the best way to solve the problems encountered. But this is not the case. In fact, the feature based on the dispersion of parking spaces is one of the most powerful predictors of parking difficulty.

in conclusion

Using Google's model, the researchers were able to generate an estimate of parking difficulty for any location and time. Below are a few examples of the system's output, which is used to provide an estimate of parking difficulty for a given destination. For example, Monday mornings are difficult for parking throughout the city, especially in the busiest financial and retail districts. On Saturday nights, it becomes busy again, but it is mainly concentrated in areas such as restaurants and attractions.

Output of the parking difficulty model for the Financial District and Union Square areas of San Francisco. Red indicates a higher confidence level of parking difficulty. Top row: 8am (left) and 9am (right) on a typical Monday. Bottom row: The same times on a typical Saturday.

As a winner of Toutiao's Qingyun Plan and Baijiahao's Bai+ Plan, the 2019 Baidu Digital Author of the Year, the Baijiahao's Most Popular Author in the Technology Field, the 2019 Sogou Technology and Culture Author, and the 2021 Baijiahao Quarterly Influential Creator, he has won many awards, including the 2013 Sohu Best Industry Media Person, the 2015 China New Media Entrepreneurship Competition Beijing Third Place, the 2015 Guangmang Experience Award, the 2015 China New Media Entrepreneurship Competition Finals Third Place, and the 2018 Baidu Dynamic Annual Powerful Celebrity.