The existing human target device-free indoor localization algorithm based on received signal strength (RSS) is difficult to give consideration to artificial workload, time consumption, and positioning accuracy under the circumstance of environment changes. In view of this problem, this paper proposes a device-free localization algorithm based on transfer clustering and fusion variational auto-encoder (FusVAE) in changing indoor environment. After environment changes, a small amount of RSS samples without labels were collected. Then, a semi-supervised fuzzy C-means clustering based on metric learning (SFCMML) was proposed to accurately cluster and label the samples, and the original model was retrained, where only a small amount of artificial work and process time was required to achieve a high localization accuracy for the original model in the new environment. In addition, aiming at the problem that the RSS samples collected in the new environment were in small quantity, the FusVAE was constructed based on coordinate fusion to generate RSS samples in the new environment for data enhancement, which could enrich the quantity and quality of the RSS samples, improve the generalization ability of the model, and enhance the positioning accuracy. Experimental results show that under the circumstance of environment changes, the average positioning accuracy of the proposed algorithm reached 88.6%. Compared with the algorithms of the same type in the same field, the proposed algorithm had higher positioning accuracy and better environmental adaptability, which is more applicable to device-free indoor localization in changing environment.