基于改进SSD的铁路障碍物检测研究

Research on railway obstacle detection based on improved SSD

铁路安全问题中由障碍物造成的安全事故占比大,后果严重。目前,铁路障碍物的检测方法存在检测精度低、实时性差的问题,并且需要耗费大量人力。为此,文中将目标检测技术应用到铁路障碍物检测中,通过实时监控预警辅助人工巡检,提高检测效率。首先选取Faster R-CNN和SSD两种目标检测算法,根据特征提取网络性能对比,选择两种性能较好的特征提取网络VGG-16和MobileNet-v2,完成4种目标检测模型的搭建;然后通过修正参数训练,对两种模型进行对比,确定最佳模型为SSD_VGG-16;最后在此基础上,从反卷积特征融合方面入手,对模型进行改进。测试结果表明:改进后的New_SSD_VGG-16模型的平均精度(mAP)为82.4%,单张图片平均检测时间为0.071 s;相较于Faster R-CNN及未改进的SSD_VGG-16模型,文中模型在保证一定检测准确率的同时,还可以提升检测速度。

In railway safety problems, safety accidents caused by obstacle account for a large proportion, and the consequences are serious. The railway obstacle detection method has the problem of low detection accuracy and poor real-time,and can take a lot of manpower. On this basis,the target detection technology is applied to railway obstacle detection,and realtime monitoring and early warning are used to assist manual inspection,so as to improve the detection efficiency. The Faster RCNN and SSD object detection algorithms are selected,and two better performance feature extraction network VGG-16 and MobileNet-v2 are selected according to the feature extraction network performance comparison, so as to complete the construction of four object detection models. Two models are compared by training the correction parameters to determine the best model as SSD_VGG-16. The model is improved from the aspect of the deconvolution features fusion. The testing results show that the mAP of the improved New_SSD_VGG-16 model is 82.4%,and the average detection time for a single image is0.071 s. In comparison with Faster R-CNN and unimproved SSD_VGG-16 model,the model in this paper can improve the detection speed on the basis of ensuring a certain detection accuracy.