Long Range (LoRa) communications are largely used in low-power wide-area networks for monitoring applications, as they enable devices to communicate over long distances. However, as the number of devices increases, collisions of LoRa frames occur frequently, which degrades both LoRa system performance and its scalability. In this paper, we propose the SF-DS scheme that decodes collided LoRa signals based on the detection of the frequency of each symbol frontier. SF-DS has several distinguishing features compared to the other protocols from the literature: it is able to reduce the detection error when symbols of similar values are superposed, does not need to store the raw signal, and does not need to perform a large number of FFTs. Although SF-DS requires modifications at the gateway side, it is fully compatible with legacy LoRaWAN communications from end-devices. Our realistic simulations under severe channel impairments and with several configurations show that SF-DS outperforms the existing collision decoding algorithms, hence increasing the throughput and thus the scalability of the network.