Spectrum sensing based on the covariance matrix algorithm does not need the prior information of the authorized user and is easy to implement. However, at low SNR, the difference between the covariance matrix elements becomes smaller, and the detection performance needs to be improved. To this end, using the similar median value to describe the accuracy of a group of data, an improved spectrum sensing algorithm based on the ratio of the maximum of the covariance matrix to the average of the similar median value is proposed. The proposed algorithm does not need the prior information and noise power of the authorized user signal, and the analytical representation of the decision threshold is derived. The simulation results under Gaussian channel shows that When SNR=-8dB, the detection probability is about 20% higher than MEAM algorithm.