Cheap and easy solution for people tracking in shopping malls, stores, office buildings and on the streets.

Around two years ago Nikita realized how hard is it for retail not to have the insights about their audience that e-commerce readily has, and not to understand something about their customers. We understood that we could change that using machine learning an computer vision. Having researched the offers on this market, we realized that we could make something cheap and, most importantly, easy to use.

We implemented a service for counting and tracking people. The user draws straight lines of interest on the picture, and the events of crossing the chosen lines are counted and remembered. These lines could be reconfigured in real time.

A video stream is taken from a Raspberry Pi or a phone's camera or any other RTSP-enabled device, then it is relayed to our intermediate server. There, videos are aggregated and prepared to be processed on a server with a GPU. Next, people are detected on the video with YOLO neural network. For each person, we try to match them with persons on adjacent frames, using their position and apperance features from a simple separate neural network. Next, the video, now enriched with markup of the people, is transmitted back.

Our service consists of two client-side apps: one for viewing the real-time statistics, and one for drawing lines of interest. Both apps are available for web and Android.

We're taking part in KONE's competition in people flow. We're good at detecting, counting and tracking people, and with this, we can, for example, optimize elevator waiting times in a building.

We're proud to have two platforms. At the beginning, we were afraid that we're going to simply run out of time and fail everything.

We learned to mock up such services and many other new stuff.

We're thinking of making a startup around this technology.

Built With

Share this project: