Synaps Labs

AI platform for targeted video advertising

WEB

AI

About the client

Synaps Labs is a unique service for traffic analysis and advertising setup to increase its effectiveness on digital billboards in the city and shopping centers.

The service's operation is based on our developed algorithm at the intersection of artificial intelligence, machine learning, and computer vision.

Task

The client came to us at the idea stage for a startup. The idea was to target outdoor advertising on video billboards to selected target segments in real-time. For example, tell all premium car owners about start of sales at an outlet or broadcast IKEA advertising for family station wagons. We immediately saw the project's huge potential and enthusiastically participated in the development.

The task before us was to create software from scratch that would help cameras track cars belonging to certain target segments and manage billboard operation based on the obtained data. All to show advertising only to the target audience and increase its effectiveness.

Solution

Developed the world's first advertising AI platform for analyzing passing cars in real-time through surveillance cameras. The system manages digital billboards on city roads and helps control ad displays, focusing only on the target audience.

The foundation was our created computer vision system, using both advanced convolutional neural network architectures and established classical machine image analysis algorithms. Cameras were produced separately with our software on the embedded NVIDIA Jetson module.

Software components were implemented in Python, Go, and C++ programming languages. For working with neural network algorithms, we used keras, pytorch, tensorflow, and TensorRT frameworks.

AI-powered cameras

Initially, we used third-party supplier cameras, which allowed launching the first real test advertising campaign in a record three months from the start of development. After analyzing the quality of existing market devices, we concluded that we needed to develop not only our own software for object detection and license plate recognition but also the computer vision cameras themselves. Experience using ready-made solutions in real conditions allowed identifying bottlenecks to then perfect our own product.

During the development of our own camera, we actively collaborated with ITMO professors and engineers with extensive practical experience in creating devices designed for all-weather operation. We had to dive deep into the topic to choose the optimal device layout.

As a result, we and the client decided that cameras would be produced separately with our software on the embedded NVIDIA Jetson module. This allowed achieving perfect image quality and solving all technical equipment complexities.

Created software for real-time analysis of car traffic

Cameras automatically track all main characteristics for fine-tuning targeted advertising: car brand, model, and year of manufacture. We also learned to enrich data from external sources in real-time.

We produced and placed 25 cameras at the most premium locations where main traffic arteries pass to cover the maximum number of cars. This allowed us to quickly gather many photos of cars of all brands and models from different angles.

After that, we wrote an interface where the car model could be precisely selected and found people to manually label a sample of photos - this is how we trained the neural network.

No display: Absence of target audience

Time: 17:42:00

Display active: High concentration of target audience

Time: 17:44:26

Our software tracks the advertiser's target audience and shows advertising only when it passes by the billboard. Total time for detection, recognition, and audience analysis takes less than 0.5 seconds. The camera also reads vehicle speeds and analyzes contact time with the advertising message.

This brought outdoor advertising effectiveness to a new level and allowed more rational budget allocation.

Designed and implemented an advertiser's personal account for better study of target audience segments and targeting

The advertiser's personal account allows planning and tracking campaign progress in real-time. Based on camera information about the audience, advertising can be very finely tuned: for example, showing it only to men aged 40-45 driving Lexus cars.

We added many useful functions so advertisers could rationally spend their advertising campaign budget. For example, comparing media metric indicators of targeted placement with parameters of classical digital outdoor advertising purchases. Also planning reach, frequency, and number of contacts with less than 10% error.

Moreover, you can remotely manage cameras, adjust focal length, and even watch in real-time which cars are currently in the advertising message's contact zone.

Ensured advertisers could optimize outdoor advertising at the same level as internet advertising

After initial results in digital advertising, advertisers always analyze them, draw conclusions, and make adjustments. For example, they might change the gender of potential buyers in settings or change the headline in the advertisement.

We wanted our advertisers to get tangible numbers from outdoor advertising too, earn more with us, and for advertising to find its buyer. Therefore, after campaign completion, we implemented automatic report generation with indicators for target audience contacts, reach, frequency, and other KPIs.

Thanks to data integration, the impact of outdoor advertising on user activity can be assessed: website visits, calls, conversion to sales.

All this allows managing billboard content, changing targeting settings according to the actual audience, and thereby increasing conversion by up to 30% with maximum reach.

Ensured data confidentiality

For us and the client, it was important to maintain complete audience anonymity and act in accordance with Russian legislation.

We taught the neural network to automatically blur the windshield and license plate in photos before saving to the database. During recognition, the service receives a hash of this number, instantly encrypts it, and saves it to the database in anonymized form.