This paper presents a novel approach that uses the passive radio frequency identification (RFID) technologies to automatically classify tagged objects in a noisy environment and detect their motion status in a real-time manner. We use the relative variance of response rate from the RFID data and propose an extended Hidden Markov Model (HMM) to annotate and an online Viterbi algorithm to decode the data sequences. An adaptive change-point detection algorithm is applied to dynamically adjust to the changing environment, based upon which to detect motion status. The performance of the approach was evaluated using several measurements including accuracy and latency under a diverse set of conditions. With appropriate configurations, high average recognition accuracy and low average latency can be achieved.