Genomic selection is revolutionizing the dairy industry. Genotyping costs have been decreasing, but large-scale genotyping remains expensive. Imputation techniques can be used to reduce genotyping costs and encourage more producers to participate in genomic selection programs. Most imputation software was developed for human applications. Contrary to humans, a large amount of family information is usually available in livestock. We have developed an imputation program (FImpute) that uses both family and population information. Family imputation is followed by population imputation based on a sliding window technique, which uses family-based reconstructed haplotypes. The software was tested in Holstein (HO), Jersey (JE) and Brown Swiss (BS) breeds to impute 50k genotypes from 3k and 6k genotypes. For HO, JE and BS, there were 67,160, 5,786 and 2,031 animals genotyped with Illumina BovineSNP50 BeadChip and 39,277, 7,462 and 305 animals genotyped with 3k with GoldenGate technology, respectively. The 50k genotypes of the youngest animals, corresponding to 20,031 (HO), 1,289 (JE) and 209 (BS) were reduced to 3k and 6k. Correct call percentage for HO, JE and BS for imputation from 3k was 97.81, 97.07 and 95.91 and from 6k was 99.47, 99.12 and 98.97, respectively. Compared to 3k, 6k resulted in better accuracy for animals and breeds with lower family information. The longest computing time was 40 minutes for HO (15 parallel jobs), considerably less than using BEAGLE and with higher overall accuracy. In conclusion, using both family and population information yielded highly accurate imputation from low-density genotypes in the three dairy breeds.