一种基于ResNet的车钩状态识别方法及其应用

A Method of Coupler State Recognition Based on ResNet and Its Application

针对传统视觉方法在车钩状态识别方面存在的应用场景单一、实时性差、准确率低的问题,文章提出一种基于ResNet的车钩状态识别方法,其通过在原图上增加高斯噪声和修改曝光率、饱和度、角度等图像参数的方法进行数据增强,使车钩状态识别准确率在测试集上达到97%;利用开源加速库TensorRT 6.0将ResNet50模型部署到边缘计算平台,所测得的图片推理时间可满足机车自动驾驶中对车钩状态实时监控的要求。

Aiming at the problem of single scene,poor real-time performance and low accuracy of traditional vision in coupler state recognition,this paper proposes a coupler state recognition method based on ResNet.By adding Gaussian noise to original image and modifying image parameters such as exposure,saturation and angle,the data is enhanced,the coupler state recognition is realized,and the accuracy of coupler state recognition reaches 97%in the test set.A ResNet50 model is deployed to an edge calculation platform by using the open source acceleration library TensorRT 6.0,and its reasoning time can meet the real-time monitoring requirement of coupler state recognition in locomotive automatic driving.