Research On Monocular Vision-based Locomotive Detection And Distance Estimation Method For Docking Operation

In order to meet the ranging requirement of locomotive docking operation in railroad transportation system,this paper systematically investigates monocular visionbased object distance estimation methods,and a theoretical framework of monocular vision distance estimation method is proposed,which involves multi-scale object detection,rail track detection,object localization and distance estimation methods based on monocular vision.The main research results are as follows:In terms of multi-scale object detection,an improved network based on yolov5 is constructed,which includes one more small-scale prediction head structure is added to enhance the learning ability of the network for small scale objects detection,and the original structure of neck-network is replaced with the weighted bi-directional feature pyramid network to balance the semantic information difference varies among difference feature layers,so as to solve the scale imbalance problem and improve the detection performance for multi-scale objects.Besides,in terms of model training,data augmentation and loss function calculation methods are studied.In terms of model inference,the weighted average non-maximum method is used to improve the accuracy of bounding box regression.In terms of monocular vision-based object localization and distance estimation,firstly,a method based on rail track detection and tracking algorithm is proposed to locate the object on the same rail track as the currently moving locomotive.Secondly,the mathematical model of monocular vision distance estimation based on geometric imaging relationship is established,and a method based on vanishing point detection is proposed to estimate the camera pitch angle.On the basis of this,an adaptive object bottom edge localization method based on image super-resolution is proposed to effectively improve the accuracy of bottom edge localization,then the object distance can be estimated based on the proposed mathematical model,and improve the accuracy of traditional distance estimation method.In terms of verifying the accuracy of the proposed monocular vision distance estimation method and the detection performance of the improved network,an experimental platform of ranging system is built,and a dataset for network evaluation is constructed.The experimental results show that the proposed network can improve the detection performance for multi-scale objects detection,especially for small scale objects detection;the proposed monocular vision-based distance estimation method can improve the ranging accuracy of traditional method and effectively reduce the ranging error at long distance.The above research results have important theoretical significance and practical value for promoting the informatization and intelligence of railroad transportation system.