This paper presents a novel, cost-effective and easy-to-deploy solution to discriminate the direction of goods crossing a UHF-RFID gate in warehouse scenario. The system is based on a grid of UHF-RFID tags deployed on the floor underneath the gate equipped with a single reader antenna. When a transpallet crosses the gate, it shadows the tags of the deployed grid differently, according to the specific direction, namely incoming or outgoing. Such distinguishable signature is employed as input of a recurrent neural network. In particular, the number of readings for each tag is aggregated within short time-windows and a sequence of binary read/missed tag data over the time is extracted. Such temporal sequences are used to train a Long Short-Term Memory neural network. Classification performance of the proposed method is shown through a set of measurements in indoor scenario.