The invention provides a retrospective off-respirator respiration gating method of a cardiac image sequence. According to the method, firstly the Laplacian eigenmap in a manifold learning method is used to carry out dimensionality reduction processing on a matrix with the storage of ECG gating cardiac image sequence data to obtain a low-dimensional coordinate matrix embed in a high-dimensional observation data point set, then the Euclidean distance between adjacent feature vectors in the low-dimensional coordinate matrix is calculated, the local maxima of the Euclidean distance is detected and is used as the selection position of a gated frame, and thus a gating image sequence with the removal of respiration motion artifact is obtained. According to the method, the matrix formed by the gray values of all pixels in an image is directly analyzed, and the respiration motion information in the cardiac image sequence is obtained. According to the method, only the solution of the feature value of a sparse matrix is needed, the manual involvement of an operator is not needed, and the method has the advantages of low computational complexity, high degree of automation, and low application cost. Furthermore, only local distance information is used in the method, and a gating result is not sensitive to noise.