SOM与HL融合的地铁异物分类算法

地铁站台门与列车门之间异物的检测和对异物的种类判别是保障乘客安全乘车和列车安全运行的基础。使用改进的Self-organizing Map(SOM)分别学习训练图片的数据分布和分类标签的数据分布,再通过Hebbian Learning(HL)学习图片SOM神经元和对应标签SOM神经元之间的数学关并编码在HL矩阵中,最后以查表方式完成异物分类。研究结果表明:改进的SOM+HL模型把分类准确率从原始模型的64.44%提高到72.6%;增加PCA(Principal components analysis)模块的SOM+HL模型使异物检测分类器的分类准确率从72.6%提高到86.2%,且其在NannoPC-T2嵌入式板上的检测速度从45FPS提高到60FPS,在满足分类精度的同时也实现了异物实时分类。且有PCA模块的模型在NannoPC-T2嵌入式板上检测速度为60FPS,移除PCA模块后其在NannoPC-T2嵌入式板上的检测速度为45FPS。

Subway Foreign Object Classification Based on SOM and HL Fusion

Detection and identification of foreign objects between metro shield doors and train doors are the basis of ensuring the safety of passengers and train running.In this paper,the improved Self-organizing Map(SOM)is used to learn data distribution of training pictures and data distribution of classification labels,and the Hebbian Learning(HL)is employed to learn the mathematical relationship between picture SOM neuron and corresponding SOM neuron,and it is decoded in HL matrix.Finally,the foreign body classification is completed by looking up tables.The results show that the improved SOM+HL model improves the classification accuracy from 64.44%to 72.6%,compared with the original model;the SOM+HL model with PCA(Principal Components Analysis)module improves the classification accuracy of foreign body detection classifier from 72.6%to 86.2%,and the detection speed on NannoPC-T2 embedded board also improves from 45 FPS to 60 FPS.Finally,the real-time classification of foreign bodies is realized while satisfying the classification accuracy.