With the help of perception technologies such as Radio Frequency Identification (RFID), information interaction will expand from communication devices to various non-communication items in daily life, and the era of ubiquitous personal communication networks may come. In an Ultra High Frequency (UHF) RFID system, however, signal collision often occurs during multiple tags identification. The collision resolution across a Media Access Control (MAC) layer and a physical layer can significantly improve communication efficiency. As a delay code, Miller code is often used in the UHF RFID system. Because its baseband signal has more subcarriers, it has higher anti-noise performance. This paper adopt the cross-layer resolution and proposes a new collision separation and decoding algorithm for the Miller code of the RFID system. This algorithm separates the collision signal by clustering and dictionary matrix firstly, and then decode the separated signal by Viterbi method. In this simulation, we generate a set of simulation data to evaluate the proposed algorithm. The simulation results show that the separation efficiency of the proposed algorithm is about 25% higher than that of traditional methods. Furthermore, we evaluate the performance of the proposed algorithm in this measured data, and obtain 8% higher system throughput than a traditional dynamic ALOHA system.