Point Cloud Semantic Segmentation Of Overhead Contact System Based On Mobile Laser Scanning

Periodic inspections and maintenances of the overhead contact system(OCS)are necessary for the railway department to keep safe railway operation.With the continuous growth of railway mileage,it is urgent to develop intelligent and efficient means for OCS inspections.Mobile laser scanning(MLS),a mobile Li DAR-based technology for rapid three-dimensional point cloud acquisition,has been applied to OCS inspections for measuring the geometric parameters of contact wires.Recognizing the OCS components from point clouds is the key to the success of measuring OCS geometric parameters with MLS technology.However,the factors,such as the complexity of OCS structures,the variety of OCS components,and the large scale of the MLS point clouds collected by OCS inspections,increase the difficulty of recognition.Nowadays,how to recognize and measure each kind of OCS components from MLS point clouds is still a critical problem of intelligent OCS inspections.This study introduces the deep learning approach to segment the MLS point cloud collected by mobile 2D Li DAR,recognizing multiple OCS components on the point level.The main work of this study can be summed up as follows.A method is proposed to extract local features of point cloud frame by frame.The first step of the method is to group points into local regions at each frame by an iterative region generation algorithm.Then the Point Net is used to extract features from the generated local regions.Finally,the extracted features are fused by the recurrent neural network(RNN)to produce local features.Due to the avoidance of calculating pointto-point relations at different frames,the proposed method can reduce the computational overhead of feature extraction.A stream data processing model and a batch data processing model is proposed in this paper.Both models utilize the proposed local feature extraction method to extract features at multiple scales.The types of RNN is different in these two models.The stream data processing model utilizes unidirectional RNNs to fused feature,supporting online point cloud semantic segmentation during data acquisition.The batch processing model uses bidirectional RNNs for feature fusion,making the information of features exact.Compare with the stream data processing model,the batch data processing model can segment point cloud with higher speed and accuracy.This study trains and tests the proposed models based on the real MLS point clouds of OSC.On the test dataset,the mean Intersection-over-Unions(m Io Us)of stream data processing model and batch data processing model are,respectively,96.12%and 97.17%.Moreover,further experiments with different configurations are conducted to verify the designs of feature fusion and multi-scale feature extraction.The results show that feature fusion and multi-scale feature extraction make markable contributions to the reduction of errors.