The limited applications of the traditional machine learning algorithms and the transfer learning algorithm are considered in this study. The improved manifold embedded distribution alignment (MEDA) algorithm is utilized to improve the detection accuracy in the cross-subject emotion recognition. The MEDA algorithm in the manifold space could reduce the data drift between domains by popular feature transformation, which can adaptively and quantitatively estimate the weights of edge distribution and conditional distribution. This article proposes an improved manifold space distribution alignment algorithm to address the problems of large feature dimension and possible bad features. An improved minimum redundancy maximum correlation algorithm is introduced for feature selection. The computational complexity is reduced, the associated features are selected, and the decision-level fusion on multiple groups of recognition results in multi-source domain is performed to further improve the transfer learning effect. The analysis results of SEED data set and the measured data set show that the distribution alignment algorithm in the manifold space is better than those of the support vector machine, transfer component analysis and joint distribution adaptation. The overall recognition accuracy is improved by 8. 97% , 4. 00% , and 2. 89% , respectively. The improved distribution alignment algorithm in manifold space has improved the recognition accuracy of each subject, and the overall recognition accuracy is improved by 3. 36% . Therefore, the effectiveness of the proposed method is verified.