In order to improve the accuracy of sparse reconstruction based on compressed video sensing and achieve a higher quality of reconstructed video frames，considering videos' sparsity features in different domains，this paper proposes a dualsparsity reconstruction algorithm based on multi-dimension reference frames (MＲF-DSＲ) in compressed video sensing． Firstly，a dualsparsity reconstruction model is proposed that video frames group sparsity and laplacian sparsity are both utilized to restrict the reconstructed videos sparsity． Besides，the concept of multidimension reference frames is elaborated in this paper，where half-pixel dimension reference frames and scaling dimension reference frames based on time dimension reference frames are introduced to obtain match-block groups with higher sparsity． Lastly，a fast diamond searching algorithm is presented to implement largescale regional searching with low complexity，which，through the coarse and fine search process，determines the position of optimal time dimension reference frame similar block，then for quick search a small scale in the same position of dimensional reference frames． Experiment results manifest that the proposed MＲF-DSＲ outperforms the state-ofthe-art compressed video sensing reconstruction algorithm both on subjective and objective criteria．