In passive RFID systems, tags are arbitrated on the Medium Access Control layer by the Framed Slotted Aloha protocol. Tags select an arbitrary slot to send their information, such that the reader receives slots with no responses, single tag responses or multiple tag responses, generating empty, singleton and collision slots, respectively. In order to maximize the system throughput, precise knowledge of the tag population competing in the read range of the reader is required. Normally, a reader does not have that information and needs to estimate it in every identification round from the statistical information collected: empty, singleton and collision slots. However, this estimate is poor since conventional readers are not capable of determining how many tags actually respond simultaneously when a collision slot occurs. In this work, we use a physical layer reader architectures to extract the exact number of tags participating in a collision. We employ this additional information and a technique that permits to successfully extract tag signals from collision slots. This establishes a much more accurate Maximum Likelihood estimator which outperforms other estimators studied in depth in previous works.