3D视觉技术在汽车轮胎字符识别中的应用

Application of 3D Vision Technology in Automobile Tire Character Recognition

基于2D视觉的轮胎字符识别方法存在效率低、精度差、易受光照条件影响等不足,从而导致系统工作不稳定。因此,提出了一种基于3D视觉的轮胎字符识别算法。首先选用3D线扫激光传感器获取轮胎胎面字符的三维点云数据,根据点云数据在Z轴上的高度特征将其转换为灰度值,然后采用以ResNet50为骨干网络的改进型DBNet算法,结合Nadam方法对DBnet算法进行训练优化。在此基础上,进一步采用模型剪枝技术,在保证算法精度的同时,压缩模型参数,提升算法速度,大幅减少了计算量。结果表明,在相关的4种检测算法下,该方法获取到的数据集的准确率高于传统方式数据集8.05%~12.2%,改进的DBNet检测算法结合CRNN识别算法后,在该方法获取的数据集中预测准确率达到了95.45%,单张图像预测速度由107 ms缩减到了45 ms,模型大小也由142.7 MB减少到了15.83 MB,为轮胎字符快速准确识别提供了一种新型的技术方案。

The tire character recognition method based on 2D vision has some shortcomings,such as low efficiency,poor accuracy and easy to be affected by illumination conditions,which leads to the system's unstable work.Therefore,a tire character recognition algorithm based on 3D vision is proposed in this paper.Firstly,3D line sweep laser sensor is used to obtain the three-dimensional point cloud data of tire tread characters,which is converted into gray value according to the height characteristics of the point cloud data on the Z-axis.Then,the improved DBNet algorithm with ResNet50 as the backbone network is used to train and optimize the DBnet algorithm combined with Nadam method.On this basis,the model pruning technology is further used to compress the model parameters,improve the algorithm speed and greatly reduce the amount of calculation while ensuring the accuracy of the algorithm.The results show that,under the four related detection algorithms,the accuracy of the data set obtained by the proposed method is 8.05%~12.2%higher than that of the traditional method.The prediction speed of single image is reduced from 107 ms to 45 ms,and the model size is also reduced from 142.7 MB to 15.83 MB.