Abstract: Current text-to-image cross-modal person retrieval algorithms overly rely on vision-language pre-trained large models to improve accuracy. This leads to issues such as large model parameter sizes and high computational requirements, making it difficult to meet the practical application requirements such as edge deployment. In light of this, this paper focuses on the lightweight aspect of vision-language pre-trained large models. It introduces a multi-modal joint distillation and complementary supervision strategy. Based on Huawei Ascend platform, the paper proposes a text-image cross-modal pedestrian retrieval algorithm based on a three-stage progressive knowledge distillation approach. Unlike the traditional two-stage approach that only performs independent distillation for each modality, this method first conducts distillation within the image and text modalities separately, and finally performs collaborative distillation for cross-modal semantic association in the shared latent space. Experimental results show that the student model has only 14.77% of the parameters of the teacher model. Compared with existing lightweight methods, it achieves effective improvements in mAP on the CUHK-PEDES, ICFG-PEDES, and RSTPReid datasets. It also realizes real-time inference performance with an inference time of approximately 23 ms for processing a single text-image pair on edge devices. This study confirms the effectiveness of the three-stage progressive knowledge distillation method and provides an effective pathway for the domestic implementation in scenarios such as personnel screening and intelligent security.