Based on the airborne LiDAR point cloud data and 66 sample plots data from field inventory of Chaocha Forest Farm in Genhe City, Inner Mongolia, we set the cold temperate primary and secondary forests as research subjects. The model training accuracy and estimating accuracy of different forest height models were compared (crown area weighted height model, arithmetic mean height model, LiDAR percentile height model), which were generated by field-measured forest height (Lorey's height, crown area weighted height, arithmetic mean height), LiDAR crown area weighted height, LiDAR arithmetic mean height and LiDAR percentile height, respectively. LiDAR crown area weighted height and LiDAR arithmetic mean height were extracted by double tangent tree crown recognition algorithm, and LiDAR percentile height was extracted from LiDAR point cloud directly. Then we investigated the applicability of the double tangent tree crown recognition algorithm in the study area, revealed the difference between Lorey's height and crown area weighted height, and determined the optimal explanatory variables and selected the optimal forest height model. Afterwards, the optimal tree height model was used to calculate the spatial distribution of forest height in the study area, which would provide reference data for subsequent studies of biomass and carbon storage. The results showed that the training accuracy and estimating accuracy of the field crown area weighted forest height model were well consistent with Lorey's height, yet slightly lower than the results of Lorey's height. And 50% percentile height was well fitted with the field-measured height (R2 = 0.869, RMSE = 1.366m for Lorey's height, and R2 = 0.839, RMSE = 1.392m for crown area weighted height). Estimating accuracy of each independently validated sample plot was all higher than 85%, with the average accuracy of 94.73% and the highest accuracy of 99.78%. The average estimating accuracy for mixed forests was 96.72%, higher than that for coniferous forests, 93.58%. On this basis, we selected the 50% percentile forest height model of Lorey's height as the optimal model and used it to calculate forest height and map its spatial distribution.