Hello. I'm Shibata, the manager of the Engineering Team at AnyMind Group.
In this article, I would like to share with you the challenges we are facing in our digital out-of-home (DOOH) product.
Launched last year, we are providing DOOH advertisements together with ANA Trading at Haneda Airport, and AnyMind is in charge of software development/operation.
Before going into the issues, I will share more about the configuration of this system and the technology used.
First of all, regarding the configuration, this platform is a one-stop system that provides everything from delivery and schedule management to the measurement of effectiveness using a camera as shown below.
Creative / Playback Schedule Creation
Users can create campaigns, check the availability of creative to be played, and set up a playlist in the management screen. An example of the playlist editing screen can be found below this. There're playable creatives on the left panel, and users can move them to the right panel to make a playlist, and currently, the playlist is set to play on repeat.
Next, get the video content and playback schedule from the created playlist and upload it to the computing unit connected to the display units via VPN. During this time, the system will execute the prepared Ansible script from the terminal for administrative operation and apply it to all target display units.
Video Playback / Data Measurement
After that, the video is played on the display unit, and this system also uses a Face Recognition system to determine the attributes of people passing by while the video (ad) is being played. This system captures various attributes in the camera’s field of view and estimates the gender, age, and where they are facing (* 1). Below is an image of the data capture on the system. You can see that data such as the identification of a face, face orientation, emotions, and age are inferred.
And there are various other programs executing inside the computing unit:
- 1. Play video based on a schedule.
- 2. The face recognition system guesses the passing people during playback, and acquires the viewing time and attribute data.
- 3. Transfer the obtained data in a text-attribute format to the server.
- 4. Decide the next video to be delivered. (see 1)
That's the flow of how advertisements are played. Also, there is no camera data stored on the computing unit or server - all captured views are disposed of immediately.
Now that we have shown the videos and understood the attributes of people viewing the ads, the last step is in making sense of this data.
Here, the data is obtained and consolidated from each computing unit, such as user attributes and the viewing time of each ad, and is aggregated and displayed across various charts.
The figure above is of the dashboard screen, and you can analyze the data grouped by parameters such as video, gender, time duration and so on.
Dashboard *in development
Now that I’ve introduced the product overall, then I’d like to share more about what was considered during the development of this DOOH system, as well as issues that must be taken into account when dealing with hardware.
Projected hardware-related troubles
With cloud services such as AWS and GCP, you can manage risks within the SLAs they provide, and these services also have redundancy mechanisms, alerts for outliers and other functions. However, if you are handling the hardware by yourself, you will need to consider all of this on your own.
- - Which hardware is the most cost-effective in terms of stability, price and performance?
- - How much redundancy should be secured?
- - Are there any problems such as exhaust heat, Wi-Fi signal strength, or fire hazards due to power supply wiring? And also have planned power outages been taken into account?
- - What metrics should be measured at what frequency, while network and PC resources are limited?
- - If the system is not working properly, what are the possible hardware problems?
- - Are stakeholders aware of the possible hardware problems and how to deal with them?
With DOOH, there is also a risk that passersby may touch the equipment. They may get their feet caught in the power cables, or perhaps the computing unit connected to the display may be stolen. For the latter case, you’ll need to think about security as well.
(By the way, our system does not allow communication between PCs, and only the mqtts protocol for measurement is open to the server. The software is Go binary, so information is obfuscated.)
Barriers to Internet communication environment
In an ideal scenario, it would be best to determine and display video ads that can be delivered in real-time, just like web ads. However, since DOOH usually displays video ads on large displays, the size of the video is inevitably large, and it is not realistic in the current 4G environment to immediately download and play a video in milliseconds.
Also, due to the location of the computing units for DOOH, there is the possibility that a wired Internet connection is not possible. In this case, there has to be wireless connection and there have to be measures in place for unstable communication and speed. For this reason, it is necessary to limit the amount of data communicated with the server and to download video material that is to be distributed to each computing unit in advance.
It's difficult to measure and process passersby through the use of cameras
As explained earlier, our system assigns probable attributes to a passerby based on images received from the camera. This process consumes a lot of computing space, and if left unattended, may affect the display and other operations. That is why tuning is important, such as suppressing the frame rate at which attributes are inferred or limiting the CPU resources used, and a strict balance needs to be found between data and performance.
In addition, it’s necessary to consider how to exclude images on baggage and billboards, and how to correctly identify passersby wearing face masks.
In the first place, the measurement of the audience is for the purpose of capturing and analyzing this for advertisers, but there is also the issue of how much advertisers will trust the data measured by a company.
Last but not least, what do you think of the DOOH advertising industry as a whole? There are some challenges raised so far that can be repeated,
but my personal feeling is as follows:
- - The required display size, setup, and what you want to display differ depending on the location, making it difficult to standardize.
- - It is difficult to acquire the ad and render the video in real-time due to communication speed limitations, which limits programmatic delivery.
- - Even if we were to create a single ad network across all DOOH locations, it would be difficult to judge whether the ad spaces are being operated properly and how many people are viewing it. Also, methods for tracking from viewing to conversion have not been established.
- - There is a high initial cost due to the procurement and setting up of hardware, and the number of locations where this can be installed is limited.
All of these issues cannot be solved overnight. As you can see, there are still many challenges in DOOH, but the DOOH advertising market is expected to continue growing just like online advertising - and our products will develop as well.
In this way, it is a business that is often difficult both as an industry and as our product, but the DOOH advertising market is expected to continue to grow in the same way as web advertising, so We would like to continue developing it. Thank you for reading this article!
(This is a blog written by AnyMind Group’s engineers. Please see here if you would like more info)
- Japanese: https://qiita.com/organizations/anymindgroup
- English: https://medium.com/anymind-group/tech-blog/home
※1：Appropriate operation is based on the sensing signage guidelines of the Digital Signage Consortium (https://digital-signage.jp/)