面向视力筛查的移动端距离测量系统校准策略研究——基于Vivo手机的实验验证

A Study on Calibration Strategies for Mobile Distance Measurement Systems for Vision Screening Applications—Experimental Verification Using Vivo Smartphone Platforms

摘要: 目的：探讨基于动态手势识别和距离测量的移动端视力筛查APP中距离测量功能的校准可行策略，为提升视力筛查的准确性与便捷性提供科学依据。方法：招募21名志愿者(7~20岁)，使用5款不同型号的Vivo智能手机在室内光线环境下，以0.2 m为梯度，在1.0 m~2.0 m的六个距离梯度下进行三次重复测量，共获取1890份有效数据。记录APP显示距离与真实物理距离并计算校准系数k。采用变异系数(CV)、过原点线性回归(决定系数R²)、双因素方差分析(ANOVA)评估测量稳定性及影响因素；计算平均绝对误差(MAE)和均方根误差(RMSE)，并结合Bland-Altman一致性分析对比统一校准与个性化校准的效果。结果：个体内测量稳定性极高，平均CV范围仅为0.51%~1.65%；APP显示距离与真实物理距离间存在极强的线性关系，平均决定系数R²高达0.9936。双因素方差分析显示，手机型号(F = 532.53, p < 0.001)与测试人员(F = 27.51, p < 0.001)对测距系数的主效应均显著。在校准效果上，采用全局统一系数校准的MAE为8.302，RMSE为10.87；而采用个性化系数校准后，MAE降至1.15，RMSE降至1.71，误差改善幅度达86.1%，Bland-Altman分析亦显示个性化校准具有极高的一致性。结论：该APP测距系统线性度良好，无需限定特定位置，可采用“单点校准”。因手机型号与个体面部特征对测距系数影响显著，不可套用统一系数，基于特定个体的单点个性化校准为最优策略。

Abstract: Objective: This paper aims to investigate feasible calibration strategies for the distance measurement function of a mobile vision screening APP based on dynamic gesture recognition and distance measurement, and to provide evidence for improving screening accuracy and practicality. Methods: Twenty-one volunteers (aged 7~20 years) participated. Measurements were performed using five Vivo smartphone models under indoor lighting at six distance intervals (1.0~2.0 m, 0.2 m increments), with triplicate readings, yielding 1890 valid data points. The calibration coefficient k was derived from the APP-displayed and true distances. Coefficient of variation (CV), regression through the origin (R²), and two-way ANOVA were used to assess stability and influencing factors. Mean absolute error (MAE) and root mean square error (RMSE) were calculated, and unified versus individualized calibration were compared via Bland-Altman analysis. Results: Intra-individual variability was minimal, with mean CV ranging from 0.51% to 1.65%. A strong linear correlation was found between APP-displayed and actual distances (mean R² = 0.9936). Two-way ANOVA showed significant main effects of smartphone model (F = 532.53, p < 0.001) and participant (F = 27.51, p < 0.001) on the coefficient. Unified calibration yielded MAE of 8.302 and RMSE of 10.87, whereas individualized calibration reduced these to 1.15 and 1.71, respectively, achieving an 86.1% error reduction. Bland-Altman analysis confirmed superior consistency for individualized calibration. Conclusion: The APP’s distance measurement system demonstrates excellent linearity, supporting a single-point calibration approach without requiring a fixed position. Given the significant effects of device model and individual facial morphology, a universal coefficient is inappropriate. Individualized single-point calibration is recommended as the optimal strategy.