In recent years, style migration has gradually become one of the core entry points for solving cross-domain Person Re-identification (Re-ID) tasks. However, most of the existing style migration methods only do a simple overall style transformation of the source domain data, which does not well motivate the network to utilize the characteristics of the target domain data. In order to effectively build a bridge between source and target domains in cross-domain Re-ID, we propose an Attention-constrained Background Transformation Framework (ABTF) based on CycleGAN, which can maximally transform the source domain data into intermediate samples that are consistent with the background of the target domain data without changing the pedestrian foreground of the source domain. Specifically, on the one hand, ABTF integrates the cycle consistency loss design idea of CycleGAN with the introduced Attention Constraint Module (ACM), which enhances the network's attention to pedestrian foreground and background. On the other hand, ABTF designs a new Constrained Loss (CL) to effectively guide the network to generate higher quality “fake” samples. A series of ablation and comparison experiments on authoritative benchmarks show that the proposed ABTF achieves competitive performance.