To address the issues of data processing latency and accuracy in bus passenger flow detection due to the neglect of edge computing, a real-time and lightweight bus passenger flow detection algorithm based on a “cloud-edge-end collaborative architecture”, bus passenger flow-DETR (BPF-DETR), is proposed. Firstly, RT-DETR-r18 is adopted as the baseline algorithm to enhance real-time processing capabilities. Secondly, a lightweight iRMB module is introduced to update ResNet-18 as the feature extraction backbone. Through the inverted residual structure, the model effectively captures long-range feature interactions of passenger targets and local feature interactions of small object targets, enhancing both lightness and accuracy while improving edge applicability. Thirdly, the SSFF module and TFE module components from the ASF architecture are integrated to construct the multi-scale feature fusion module (MSFM), further enhancing the detection accuracy under multi-scale and complex environments. Finally, to validate the effectiveness of the algorithm, a ROI-based image stitching method is used to enhance the representativeness and diversity of the dataset, constructing a bus passenger monitoring dataset for training and validation. Experimental results show that the mAP@0.5 of the BPF-DETR is 96.4 %, and the model size is 32.6 MB, which are both superior to current mainstream YOLO series models. In comparison with the baseline algorithm, the mAP@0.5 is improved by 1.1 %, and the model size decreases by 16 %, meeting the accuracy and lightweight deployment requirements for bus passenger flow detection.