Aiming at the problem of signal detection in small sample and nonhomogeneous clutter, a matrix information geometric detector based on manifold filter was proposed. The signal detection problem was transformed into a geometric problem on the matrix manifold. The correlation of each sample data was modeled as a Toeplitz positive definite matrix. On the basis, each matrix was replaced by a weighted smoothing filter of its surrounding matrices. The weighted smoothing filter removed part of the clutter energy and improved the discrimination between the target and the clutter. The geometric mean of a set of secondary sample data was calculated. By comparing the distance between the matrix under test and the geometric mean matrix with the detection threshold, the signal detection was realized. Simulation results demonstrate the superiority of the detection performance of the proposed method in small sample and nonhomogeneous clutter compared with adaptive matched filtering.