The last decade has witnessed a progressive interest shown by the community on inferring the presence of people from changes in the signals exchanged by deployed wireless devices. This non-invasive approach finds its rationale in manifold applications where the provision of counting devices to the people expected to traverse the scenario at hand is not affordable nor viable in the practical sense, such as intrusion detection in critical infrastructures. A trend in the literature has focused on modeling this paradigm as a supervised learning problem: a dataset with WiFi traces and their associated number of people is assumed to be available a priori, which permits to learn the pattern between traces and the number of people by a supervised learning algorithm. This paper advances over the state of the art by proposing a novel convolutional neural network that infers such a pattern over space (frequency) and time by rearranging the received I/Q information as a three-dimensional tensor. The proposed layered architecture incorporates further processing elements for a better generalization capability of the overall model. Results are obtained over real WiFi traces and compared to those recently reported over the same dataset for shallow learning models. The superior performance shown by the model proposed in this work paves the way towards exploring the applicability of the latest advances in Deep Learning to this specific case study.