A Trained-once Crowd Counting Method Using Differential WiFi Channel State Information

This paper focuses on the problem of providing a rough count of the number of people in a room using passive WiFi Channel State Information (CSI) measurements taken by a single commodity receiver. The feature which mainly distinguishes our work from others is the attempt to emerge with an approach which does not require any dedicated training inside the specific environment where the system is deployed. Our proposal stems from the intuitive observation that features which account for em variations of CSI are expected to be less sensitive to the surrounding environment as opposed to features which account for absolute CSI measurements. We turn such intuition into a concrete proposal, by suitably identifying a set of differential CSI feature candidates, and by selecting the (two) most effective ones via minimization of the summation of the Davies-Bouldin indexes. We preliminary assess the effectiveness of the proposed approach by training once for all the system in a room, and testing the system in two em different rooms having different size and furniture, and involving people freely moving in the rooms with no a-priori movement constraints.