AI video analytics to support people counting with RapidLab
About the customer and the challenge
RapidLab focuses on creating intelligent electronics and building rapid prototypes of Internet of Things (IoT) projects, solving challenges in many industries.
During the Covid-19 pandemic many state institutions, public transportation facilities but also businesses of all sizes, were faced with the unprecedented challenge of maintaining the principles of social distance and monitoring the number of people in buildings and premises at once.
In many cases, installing physical counting gates was too expensive or simply not feasible due to the architecture of the buildings. Manual counting of people, on the other hand, due to the existence of several entrances and exits of buildings, remained a fiction.
Solution
Together with the IoT engineers from RapidLab we have built a small device allowing for automatic counting of people in facilities and vehicles, taking care of the safety of the passengers, employees, and customers. Sensors are counting each person entering or leaving, no matter if they use the same or different exit. The solution has been provided in two versions, dependent on the characteristics of the facility, to deliver the best quality of results – based on the time-of-flight sensors and video cameras. The first version relies on the application of AI on time-series data, whereas the second focuses on applying computer vision deep learning techniques to build an intelligent video analytics solution.
The project implementation included i.e. definition of data collection protocols, data exploration, investigating and building quick prototypes using different neural network architectures as well as classic machine learning approaches.
An important part of the project implementation, influencing heavily the most suitable AI architectures, has been also the optimization of the AI models to run directly on the embedded devices. Such implementation, called an Edge AI solution, processes the data without any need to send them to the cloud. This deployment method allows for achieving real-time performance without delays connected with the data transfer, independence from the Internet connection stability as well as without causing any data privacy risks.
Working with the spoken language of different quality, interruptions and colloquialisms pose challenges both for speech to text and further data processing like generating summaries. An additional challenge was caused by the fact that we were working with the Polish language, which is having much smaller support than e.g. English or German. We were able to overcome the mentioned challenges thanks to building advanced postprocessing methods for the refinement of speech to text part, as well as training our own AI models to work with the Polish language.
Contact us
How can we help you? The fields marked as required below will help us to process your request and forward it accordingly. You can expect a response within one to two business days.