In the big data environment,the number of data samples for the human detection increases rapidly. There is a great difference between sharpness and discrimination information in these data samples,so the data cannot be used directly. The traditional human detection methods based on transfer learning are suitable for the situations of no target domain sample or few domain samples only. In view of the above problems,a performance improving method based on transfer learning theory for human detection is proposed. The characteristics of classifier are utilized to calculate the similarity between source samples and target samples according to the idea of transfer learning. The selection of target samples is executed to update the classifier ac⁃cording to sample distribution. Compared with the existing methods,this method makes full use of the data,and improves the detection performance without addition of more time.%大数据环境下，可用于人体检测的数据样本数量迅速增长。这些数据样本在清晰度以及所包含的判别信息等方面有较大差别，导致这些数据无法直接使用。传统基于迁移学习的人体检测方法主要针对没有目标域样本或者目标域样本很少的情况，无法充分利用大量的数据样本。针对这一问题，提出基于迁移学习的人体检测性能提升方法，该方法根据迁移学习的思想，利用分类器的特性计算源样本与目标样本间的相似性并根据样本分布图，筛选目标样本更新分类器。相对于已有方法，该方法充分利用了数据，且在不增加检测时间的基础上对检测性能有一定的提升。
展开▼
