Our sensing software is powered by ‘deep-learning’ neural network models that processes images and audio in real-time. Features of interest are analysed ‘on the fly’ and this results in anonymised ‘meta-data’ (anonymised attributes) that are used to generate metrics of the aggregate behavioural response of humans interacting in the physical space where the sensor is deployed. No video images are recorded, transmitted or retained anywhere in the system. In technical terms, each image frame that enters into the volatile memory of our sensor processing unit is discarded immediately after analysis. Because this image frame is held in volatile memory for such a short time (typically a fraction of a second), it’s considered to be ‘ephemeral’, which means that no personally identifiable data remains anywhere in the end-to-end system.