Athletes who attended the VnExpress Marathon Quy Nhon 2020 last weekend can search for their pictures by typing their bib numbers on VnExpress official website, instead of browsing through all of the pictures in photo collections as before.

Previously, AI was applied at some large-scale running events. However, at VnExpress Marathon Quy Nhon 2020, the technology is designed to serve a different purpose, which is to detect and cluster faces in images within a few seconds.

Bib is the yellow and white sheet on athletes’ chests at the running event in Quy Nhon in 2020. Photo: VnExpress.

Athletes wear their names as well as their numbers on a sheet of thin plastic that stick to clothing called bibs. It is quite hard to scan and read the numbers on the athletes’ bibs. Bibs appear in various pictures taken during the run. The requirement for VnExpress Marathon’s technical team is to read all the bib numbers in an image with the accuracy of over 90%, regardless of tilted, distorted or partially obscured images, within 5 to 10 seconds. In addition, the team has to ensure the system stable and upgraded in case the number of images increases sharply.

“As we received the request, we immediately considered to use AI to handle those tasks,” said Nguyen Loc Vu, CTO of VnExpress.

When it comes to bib number detection, there are three challenges facing the team, which are finding the right position of the bibs, converting the image into text and reading numbers in the text.

In response, the team built an AI-powered system to recognize and segment athletes in images. The bibs will be in a specific cluster of athletes with almost 100% probability. Therefore, the system do not have to handle excess data, such as billboards, dashboard or license plates on the run, making it stand out from other existing detection systems.

Once bibs are located, character recognition algorithms digitize the information written on the bib and put it into the form of a series of letters and numbers. Then, the system filters out unnecessary characters to get the athletes’ numbers.

After that, the process of reading bib numbers is “packaged” into Docker. Each Docker can process images completely and independently, so to accelerate image filtering process or increase the number of images, administrators just need to add Docker to the system.

On average, each Docker takes about 5 seconds to read all the bibs in an image. It takes almost 8 hours to read bibs 60 thousand images by using 10 Docker. Vu said the real processed time is even shorter because images are continuously uploaded to the system during the running event. Therefore, athletes can get their photos on the same day.

Vo Trong Tan, a runner from Quang Nam, appreciated VnExpress Marathon Quy Nhon 2020 as one of the races that published photos early. It just takes one day for participants to get their photos, while it takes a week, or even a month in other races.

AI recognizes faces and bibs in images. Photo: VnExpress.

VnExpress’s racing bib number recognition system also receives good feedback for its cability of detecting the bib that is partially obscured. Tan shared that initially, he was worried the orange bid was easily blurred due to strong sunlight and he could not get any photos at a distance of 21 km. However, he was “surprised” when he found his photos easily.

The ability to read data on images depends on the quality of original image. According to CTO Loc Vu, VnExpress’s system can process all pixels of the image, so it can detect some details that are difficult to see. In addition, the system can process images at multiple resolutions to accommodate different types and lenses of the photographer at VnExpress Marathon.

After one day, there were tens of thousands of visits to the image collection of VnExpress Marathon Quy Nhon 2020. The ability to detect bib numbers with the accuracy of 93% was highly appreciated by athletes.

After 4 months of developing and testing, the system currently achieves initial success. “We will continue to complete our racing bib number recognition system. Our goal is to improve processing speed as well as accuracy to serve VnExpress Marathon events later or to provide the system to other organizations,” said Vu.

Launched on February 26, 2001, VnExpress, currently operated by FPT Online, has became the most-read online newspaper in Vietnam.

Thao Chi

Related posts: