fbpx
Connect with us

Economy & Tech

How Useful is Artificial Intelligence in Traffic Surveys?

Artificial intelligence offers a potential solution to Japan's shrinking workforce in traffic surveying. But is it good enough to replace human workers?

Published

on

A research group led by Professor Ryuichi Imai of Hosei University has studied the effectiveness of cameras and artificial intelligence (AI) in examining traffic volume. Specifically, they examined its effectiveness in the Road Traffic Census conducted by the Ministry of Land, Infrastructure, Transport and Tourism. 

The team found that it could not accurately measure the traffic volume due to its failure to recognize vehicles. However, the researchers identified three challenges and solutions for practical implementation.

Their research paper was published in the Journal of Digital Life. It is a multidisciplinary, peer-reviewed, and open-access online journal based in Japan.

The Road Traffic Census is conducted approximately every five years. It consists of two parts: a general traffic volume survey, which examines road usage, and a survey which examines the starting point and destination of traveling vehicles. 

Will AI Take Over?

In 2021, the Ministry announced its plan to replace manual observation in the general traffic volume survey with AI. This approach utilizes video images from CCTV cameras, commonly employed for road management. The shift toward AI analysis has generated considerable interest. But there are concerns that AI could take over the part-time jobs of surveyors.

Nevertheless, a survey conducted in 2021 by the National Institute for Land and Infrastructure Management (NILIM) revealed the need for more human surveyors in rural areas for traffic surveys. Therefore, it appears that AI has not completely taken over yet. 

An additional issue identified with the AI system was the decreased accuracy of the computer's recognition capabilities at night and during traffic congestion. Particularly, the accuracy between 8 pm and 10 pm plummeted to 20-30% of that of visual observation. 

Consequently, in the 2021 survey, an estimation was made for the traffic volume during the hours when AI was unreliable. This relied on past surveys conducted in similar traffic circumstances and locations where nighttime monitoring was available.

A Solution to the Declining Workforce

As the workforce decreases, the recruitment of surveyors is expected to become more challenging. Consequently, efforts are underway to develop systems that can enhance efficiency.

For instance, Imai and his research group aimed to automate two types of measurements. The first is "cross-sectional traffic volume," which summarizes the number, speed, and type of inbound and outbound vehicles. The second is "intersection traffic volume," which examines the movement of vehicles turning right, left, or going straight ahead at intersections.

Through this, they developed a technology to solve three key problems: inaccuracy in counting the number of vehicles passing through, misrecognition of vehicle type classification, and reduced accuracy at night.

Vehicle Counts

The accuracy of vehicle counts has been a challenge in existing AI methods. They often fail to detect vehicles in the vicinity of the survey point, leading to inaccurate counts. To address this, the research group utilized YOLOv3 (You Only Look Once), an AI model that detects objects in video images. It can identify vehicles with similar characteristics appearing in frames before and after they passed the survey point (each still image of the video). Therefore, the system ensured that even if a vehicle was missed at a specific "point," it would still be detected along the "line" of its trajectory. 

Because existing methods tended to misclassify vehicles that were similar in color and shape, researchers used VGG19, an AI technology that recognizes images, to improve accuracy.

The technology was tested using a 15-minute video of two one-way lanes in Osaka City. While the number of vehicles was accurately counted, the classification of vehicle types was still an issue. Sedans and trucks were classified correctly, but vans and small trucks were misclassified as larger vehicles.

Vehicle Classification

To address this issue, the research group conducted interviews with 10 experienced surveyors. They had several years to several decades of experience in traffic volume surveys. The researchers discovered that vehicle classification relied not only on the vehicle's shape but also on other features. These include the front-to-side ratio, the proportion of the windshield and other parts in relation to the entire vehicle, and the color of the license plate. Building upon this insight, the researchers developed a method that utilized specific parts of the vehicle as an identifier.

Additionally, the researchers made computer-generated graphics of various vehicles. Then, they assigned colors to different parts of the vehicle such as red for the front, green for the top side, and blue for the windshield. These graphics were used to train the AI system to learn the characteristics of each vehicle type. After repeating the previous experiments, the AI successfully classified small sedans, vans, buses, and trucks. It also maintained accurate counts within a "10% error margin," which is essential for practical application.

Nighttime Accuracy

In an effort to address the issue of inaccuracy during nighttime hours, the research group explored the use of AI technology to convert nighttime images into "daytime" images. This method was tested in conjunction with the aforementioned classification technique that employed vehicle parts as identifiers. The results showed that the accuracy of vehicle detection in the converted nighttime images (from 7 pm to 7 am) was comparable to that of images captured at the same survey point during the daytime.

During the nighttime, the luminous intensity was measured at approximately 10 lux, akin to the level of illumination in a movie theater during an intermission. Without the image conversion, nighttime accuracy was notably reduced, particularly for larger vehicles. Therefore, the conversion technology significantly improved accuracy in these instances. However, the team also observed that the conversion of nighttime images presented challenges when the vehicles were black, leading to the misclassification of vehicle types.

Based on the experiments' results, the team concluded that the three technologies employed were beneficial for traffic surveys. The research group will review existing studies and develop new methods to further enhance the accuracy of vehicle classification and explore different applications of the technology.

The research team members included:

  • Professor Ryuichi Imai of Hosei University
  • Associate Professor Daisuke Kamiya of the University of the Ryukyus
  • Assistant Professor Yuhei Yamamoto of Kansai University
  • Professor Shigenori Tanaka of Kansai University
  • Koki Nakahata, a student at the Kansai University Graduate School
  • Associate Professor Wenyuan Jiang of Osaka Sangyo University
  • Lecturer Masaya Nakahara of Osaka Electro-Communication University.


RELATED:


This article was first published on Sankei Biz by the Journal of Digital Life. You can also read the article in Japanese.

Author: Taketoshi Noma