One of the biggest challenges in CV is counting people on videos. This due to the difficulties of tracking the same person every frame. Ladorian is a targeted marketing company that has a service to show targeted ads in TVs or screen walls, they were interested in counting people passing in front and looking at the screens. I’ve decided to solve counting people in a new way, neither by tracking them but using face recognition. From a simple face detection we obtained the embedded vector of that persons and also on the same infer age and gender. It works much more accurate than people tracking. It is nearly perfect after we take off our covid masks.
The pipeline that I build works better than other commercial solutions due to the use of face-detection we’re not counting the same person many times. This is the key to the success of that project. This pipeline runs to nearly 10 FPS in a CPU providing to the adserver statistics and quantitative data. We are also adding plate recognition for petrol stations and car brand recognition to deliver more features.
Technology stack Python + InsightFace with ONNXRuntime + OpenCV
