Analysing human body shapes requires a substantial amount of anthropometric data. However, traditional manual measurement methods are often inefficient, while 3D scanning devices are expensive and inconvenient. To address these challenges, this study presents a method based on the U2-Net neural network model for extracting human body contours in complex backgrounds, feature point extraction, and circumference fitting analysis. Using data enhancement techniques, we utilized a dataset comprising 2 560 frontal and lateral images of individuals against various backgrounds, which was augmented to 42 800 images. Subsequently, a deep learning network model was trained to accurately fit chest, waist, and hip circumference measurements. Finally, 20 samples were selected for validation, and the predicted values were compared with manually measured values, showing that the errors fall within an acceptable range. The effectiveness and accuracy of this method have been validated, providing a practical solution for anthropometric data collection and body shape research in remote areas.