Abstract: The multi-dimensional detection of primary user is one of the essential problems in cognitive radio. The conventional detection shows only the spectral occupancy. Here we propose a method based on auto-regressive(AR) spectral estimation to go around the frequency-time detection for primary users in low singal-to-noise ratio(SNR). The proposed method uses AR spectral estimation to estimate the Range-Doppler spectrum of received signal, and then considers the peak of Range-Doppler spectrum as the test statistic. The expressions of the detection probability and false probability are deduced respectively. The simulation by MATLAB shows the spectral occupancy in the frequency-time dimension, receiver operating characteristic(ROC) curve. It confirms the AR model spectrum sensing has better performance than the energy detection in cognitive radio. Meanwhile, we also research the optimal order of AR model.