Software Development for Calculating Foot Traffic Using LiDAR

These two types are effective and ineffective billboards. Hundreds of people pass by the first type every minute, while the second type is unnoticed by anyone.

It's easy to distinguish between them if the marketer ordering out-of-home advertising knows the city well. But what if they live in Madrid and are placing ads in Stockholm? This is precisely why an independent tool for evaluating foot traffic near billboards is needed for such situations.

Technically, our task could have been solved easily: we install a camera, connect an AI computer vision model, and it counts people. However, our solution needed to operate in the European market, where GDPR, the law on personal data protection, applies.

Recording people on the streets for marketing purposes is prohibited by law. Our solution had to only capture information about the number of people. Therefore, a regular camera was not suitable.

To address the problem, we considered two alternatives to replace a regular camera: LiDARs and thermal imagers. Both devices can count the number of people without capturing any personal information.

LiDARs proved to be the more reliable technology. We proceeded with calculations: which LiDAR we needed, its field of view angle, and the optimal positioning near the billboard.

The optimal model turned out to be the MID-360. It's compact, has high build quality, offers high resolution, and most importantly, operates at 10 frames per second.

LiDAR transmits data in the form of coordinate sets, which are in a text file format unsuitable for any computer vision models.

Therefore, we developed specialized software that converts these coordinates into images. These images do not reveal identifiable human faces. Capturing data via LiDAR does not violate GDPR regulations

To collect data and transmit it to the server, we:

• Connected the LiDAR to a Raspberry Pi mini-computer
• Connected a 4G modem to the Raspberry Pi
• Created scripts in Ansible to simultaneously manage the devices

As our computer vision model, we chose YOLO8 because it is resource-efficient and supported by numerous open-source libraries. To train the neural network, we captured 8000 LiDAR snapshots, converted them into images, and annotated human figures.

• The LiDAR captures pedestrian figures for several minutes
• Ansible stops data capture on all devices/li>
• Our software converts LiDAR coordinates into images
• YOLO8 counts the number of people passing by the billboard during the capture period
• Raspberry Pi sends the pedestrian flow data to the server
• Marketers use this data to analyze the effectiveness of the billboar

We have prepared an engineering solution that our client and their clients, corporate marketers, can already use. Looking ahead, our future plans involve integrating the LiDAR, mini-computer, and modem into a single device capable of operating in any weather conditions.