针对传统路面平整度检测中异物干扰导致的误判问题,本文提出了一种融合加速度传感与视觉识别的多模态动态评估方法。通过时间戳对齐加速度数据与视频帧,计算竖直加速度均方根作为路面平整度指数,采用图像分类模型排除异物导致的加速度波动干扰,修正了异物导致的加速度数据波动,并将修正前后的平整度曲线在软件中展示。实验结果表明,在固定路段检测中多模态方法将检测准确率从单模态的67.9%提升至95.0%,绝对准确率提高27.1个百分点;由路面异物干扰导致RMSVA波动情况244次,成功识别并修正路面平整度测量结果206次,该方法能避免84.4%由异物导致的路面平整度测量的问题,修正后各路段竖直加速度均方根降幅达19.1%~25.0%。该方法通过“振动感知–视觉校验–动态修复”闭环处理机制,在保持检测实时性的同时,降低设备成本约65%,为道路养护提供了高性价比的动态检测方案。To address the misjudgment caused by foreign object interference in traditional pavement smoothness detection, this study proposes a multimodal dynamic evaluation method integrating acceleration sensing and visual recognition. By synchronizing acceleration data with video frames through timestamp alignment, the root mean square of vertical acceleration (RMSVA) is calculated as the pavement smoothness index. An image classification model is employed to eliminate acceleration fluctuations induced by foreign objects, and correct acceleration data fluctuations caused by foreign objects. The pre- and post-correction smoothness profiles are visualized in the software interface. Experimental results demonstrate that the multimodal approach in the fixed road section detection improves detection accuracy from 67.9% to 95.0% in single modality, achieving an absolute accuracy increase of 27.1 percentage points. RMSVA fluctuation events caused by the interference of foreign objects on the road surface were found to
