In order to solve the problem of blind detection of MIMO system, this paper takes maximum-likelihood sequence detection as the criterion and derives the formulas to get a model based on SemidDefinite Relaxation. The rank of SDR solution equals to the number of the transmit antennas. For the rank of SDR solution is greater than 1, a new method is proposed to approximate the solution of the original problem, which combines the eigenvector approximation method and randomization method. By setting the upper limit of objective function, the proposed method could judge the number of detection sequence adaptively and reduce constrains number and the number of solving SDR. The analysis shows that the computation complexity of proposed method has linear relationship with the number of transmit antennas. At last, simulation results indicate that compared with Rank-1 algorithm, the proposed detector could provide the same bit error performance with decrease of computation cost, and validate the linear relationship between the computation complexity and the number of transmit antennas.