Abstract
Real-time obstacle detection on railway tracks is crucial for preventing accidents. Current solutions typically rely on static sensors monitoring limited zones and black-box methods with unpredictable behaviour for different scenarios. In our work, we present an interpretable and real-time railway track and obstacle detection method using an on-board LiDAR (Light Detection And Ranging) sensor. Our method comprises two main steps: point cloud-based railway track detection and obstacle detection for the predicted trajectory. To estimate railway tracks we define a pillar max height-based filtering, followed by a point aggregation operation and a polynomial curve fitting. For the given railway tracks, we base our object detection on point clustering. Both railway track and obstacle detections are tracked in time for more robust and smooth results. To address the lack of available datasets, we create a synthetic dataset for evaluating obstacle detection. We evaluate our method on both real and synthetic data, and the results demonstrate its effectiveness in accurately detecting railway tracks and objects.
Aitor Iglesias
PhD Student
My research focuses on enhancing the reliability of Autonomous Driving functions using Machine Learning techniques.
Nerea Aranjuelo Ansa
Former PhD student
My research focuses on machine learning and computer vision for multimodal perception systems.