Object-Level Data Augmentation for Deep Learning-Based Obstacle Detection in Railways

Abstract

This paper presents a novel method for generation of synthetic images of obstacles on and near rail tracks over long-range distances. The main goal is to augment the dataset for autonomous obstacle detection (OD) in railways, by inclusion of synthetic images that reflect the specific need for long-range OD in rail transport. The presented method includes a novel deep learning (DL)-based rail track detection that enables context- and scale-aware obstacle-level data augmentation. The augmented dataset is used for retraining of a state-of-the-art CNN for object detection. The evaluation results demonstrate significant improvement of detection of distant objects by augmentation of training dataset with synthetic images.

Local and Global Information in Obstacle Detection on Railway Tracks

“…The pasting process is inspired by Fishyscapes [12], an object-enhanced dataset for measuring segmentation blind spots in traffic images, and involves image border smoothing, brightness correction, motion blur, depth blur, and Gaussian noise. Similar augmentations have also been proposed by [34], [37], [38], [56] for use in railway anomaly detection, as there are no public real-world railway obstacle detection datasets. Our synthetic obstacle testing method is also in line with the one proposed by Boussik et al [14], who instead use a GAN to blend images of railways and obstacles on the RailSem19 dataset.…”

“…Uribe et al [37] follow a similar approach with the same shortcomings. Learning-based object detection methods have also shown some success in detecting humans, trains, or luggage [4]- [8], [38]- [41], but they rely on custom datasets and are limited to a fixed set of object categories.…”

Toward the Enhancement of Rail Sustainability: Demonstration of a Holistic Approach to Obstacle Detection in Operational Railway Environments

“…As explained in the following, these activities involved obtaining permissions for demonstration, the selection of a demonstration site, preparation and setup, scenario design, a live demonstration of the data recording and data processing for the purpose of obstacle detection, and distance calculation. Detailed descriptions of the SMART2 solutions for obstacle detection and distance calculation can be found in a number of publications such as [10,11]. This paper focuses on the results of the demonstration of this SOLUTIONS, and on the analysis of the OD system performance and limitations based on the obtained demonstration results.…”