Spectrum sensing is a fundamental functionality for cognitive radio networks to detect spectrum resource status and to provide the opportunity for the cognitive user to use the under-utilized frequency resource without causing harmful interference to primary user. Energy detection, which is the most widely used in Cognitive Radio system, belongs to a kind of crude detection method and is not sensitive to noise uncertainty, especially under low SNR condition, detection performance will decline. On the other hand, covariance detection is more feasible to avoid the uncertainty of noise effect on detection performance. In this paper, the fractional-step detection based on energy and covariance detection is proposed to achieve spectrum sensing. This algorithm combines the simplicity of energy detection and the fine statistical function of covariance detection, and improves the accuracy by detecting step by step. But it will cause more hardware overhead. So to use fractional-step algorithm in low SNR has to compromise in the detection accuracy and the complexity of the algorithm.