Key Components Recognition Of Locomotive Running Gear Beased On Three Dimensional Measurement

Anomaly detection of locomotive running gears is the focus of routine maintenance of railways and an important part of railway safety precautions.At present,the safety inspection of locomotive running gears in China has gradually shifted from the traditional manual detection to the automatic detection based on 2D machine vision.However,misdetection and missed detection has always been difficult to avoid.In recent years,laser 3D measurement technology has developed rapidly.3D data can record clearer and richer information of the measurement object,and it has the advantage that traditional 2D images can’t be compared.This makes machine vision systems based on 3D data have the potential to fully surpass 2D machine vision.Therefore,we attempted to build a key components identification system based on laser 3D measurement technology for the locomotive running gear,which is used as the preliminary work for the new automatic detection system of locomotive running gear.The main work of this thesis is as follows:1.Because of the actual measurement environment of the track side is very complex,data acquired by the laser 3D scanner contains a large amount of noise.The commonly used methods based on bounding boxes or spatial neighbor statistics are difficult to completely filter out this type of noise.We expanded the principle of median filtering in 2D image processing to 3D,and used the k-neighborhood instead of the spatial neighborhood to realize the k-neighborhood adaptive median filtering of point clouds.Experiments show that the point cloud filtering method in this thesis can effectively remove the noise in the original 3D data,and the filtered data is sufficient to meet the needs of subsequent detection.2.The original 3D data collected by the laser 3D scanner contains too much redundant information,which causes many obstacles to the subsequent processes of data storage,transmission,and analysis and processing.Aiming at the characteristics of target data,we designed and implemented a fast simplification algorithm that uses two-level non-uniform blocks to simplify the running gear 3D data.Experiments show that compared with the traditional non-uniform grid method,the method in this thesis can achieve a higher reduction rate on the premise of equal accuracy;compared with the first-level non-uniform grid method using Gauss map,this method achieves a similar accuracy while the computational efficiency has been greatly improved.3.Due to the obstruction of the protruding parts in locomotive running gears,a blind zone will appear in the 3D data.Therefore,it is necessary to use a multi-viewpoint registration method to obtain a complete 3D appearance of the running gear.This thesis implements a variety of registration algorithms,and compares the accuracy and computational efficiency.Experiments show that with pre-existing features and reduced data,the simplified data-guided registration method is the fastest,and the registration method based on feature matching has the highest accuracy.4.This thesis designs and implements a key part recognition and location method based on 3D feature extraction and template matching.First,extract the key points in the template and the target data,and use FPFH(Fast Point Feature Histograms)to describe the neighborhood features of the key points;then,based on the Euclidean distance between the FPFH descriptors,the matching points are searched among the key points in target point cloud,and the set of matching points with weights are obtained;finally,the key components are identified and positioned using cluster analysis in the set of matching points.Experiments using bolts as an example have shown that this method can roughly locate the position of the target component,but the accuracy is low,and misidentification is difficult to avoid.Therefore,this thesis improves the recognition algorithm for bolts.The Hough transform method is used to establish a strict classifier for the pre-recognition results.Experiments have shown that this improved method can effectively and reliably identify the presence and exact location of bolts.