Conditional average treatment effect (CATE) is designed to capture the heterogeneity of treatment effect across subpopulations. In this paper, we propose a new nonparametric estimation strategy for CATE based on the propensity score and projection theory. The proposed approach has two advantages over the existing ones. First, it does not need to estimate the two nonparametric regression functions of the outcome on many covariates for treated and control groups, and obtain their predicted values via extrapolation. Second, the proposed method includes propensity score as a new covariate in nonparametric regression model, thus it can effectively overcome the hazardous impact due to extreme weights (propensity score close to 0 or 1) in weighting estimators. Meanwhile, the proposed procedure does not rely on outcome model speci cation. We establish the consistency of the proposed estimator, and further show that it asymptotically follows an normal distribution and the associated variance can be estimated. Simulation studies indicate that the proposed procedures outperform competing ones. We further illustrate the proposed procedures by an empirical analysis of a real-world dataset.