A Deep Learning Based Approach to Object Recognition from LiDAR Data Along Swedish Railroads

Abstract: Malfunction in the overhead contact line system is a common cause of disturbances in the train traffic in Sweden. Due to the preventive methods being inefficient, the Swedish Transport Administration has stated the need to develop the railroad maintenance services and has identified Artificial Intelligence (AI) as an important tool for this undertaking.  Light Detection and Ranging (LiDAR) is a remote sensing technology that has been gaining popularity in recent years due to its high ranging accuracy and decreasing data acquisition cost. LiDAR is commonly used within the railroad industry and companies such as WSP collects large amount of data through LiDAR measurements every year. There is currently no reliable fully automatic method to process the point cloud data structure. Several studies propose innovative methods based on traditional machine learning to extract railroad system components from point clouds and have been able to do so with good results. However, these methods have limited applicability in real world problems, as they build upon hand-crafted features based on previous knowledge of the data on which they are applied. Deep learning technology may be a better alternative for the task as it does not require the same amount of human interaction for feature engineering and knowledge about the data in advance.  This thesis investigates if contact line poles can be recognized from LiDAR data with the use of the neural network architecture DGCNN. Data from two Swedish railroad lines, Saltsjöbanan and Roslagsbanan, provided by WSP was used. Point labels were predicted through semantic segmentation from which objects were distinguished using the clustering algorithm DBSCAN. The network was trained and validated on Saltsjöbanan using k-fold cross-validation and was later tested on Roslagsbanan to simulate the application of trained models on an unknown dataset. On point level the network achieved an estimated precision of 0.87 and a recall of 0.89 on the data from Saltsjöbanan and an estimated precision of 0.92 and recall of 0.83 on the data from Roslagsbanan. In the object recognition task, the approach achieved an average precision of 0.93 and recall of 0.998 on the data from Saltsjöbanan and on the data from Roslagsbanan, an average precision of 0.96 and a recall of 1 was achieved, indicating that it is possible to apply this method on railroad segments other than the one the network was trained on. Despite not being accurate or reliable enough on point level to be used for thorough inspection of the contact line system, this approach has various applications in terms of object recognition along Swedish railroads. Future research should investigate how adding additional classes beyond contact line poles would affect the results and what changes can be done to the parameters to optimize the performance. A side-by-side comparison with the current methods and traditional machine learning-based methods would be valuable as well.