Aiming at the low classification accuracy problem of large-scale Support Vector Machine (SVM) after random-projection-based feature dimensionality reduction, Linear kernel SVM based on dual random projection (drp-LSVM) for large-scale classification problems was proposed with the introduction of the dual recovery theory. Firstly, the relevant geometric properties of drp-LSVM were analyzed and demonstrated. It's proved that, with maintaining the similar geometric advantages of Linear kernel SVM based on dual random projection (rp-LSVM), the divided hyperplane of drp-LSVM was more close to the primitive classifier trained by complete data. Then, in view of the fast solution to drp-LSVM, the traditional Sequential Minimal Optimization (SMO) algorithm was improved and the drp-LSVM classifier based on improved SMO algorithm was completed. Finally, the experimental results show that, drp-LSVM inherits the advantages of rp-LSVM, reduces classification error, improves training accuracy, and all its performance indexes are more close to the classifier trained by primitive data; the classifier designed based on the improved SMO algorithm can reduce memory consumption and achieve higher training accuracy.