Railway Obstacle Detection Using Unsupervised Learning: An Exploratory Study

Abstract

Autonomous Driving (AD) systems are heavily reliant on supervised models. In these approaches, a model is trained to detect only a predefined number of obstacles. However, for applications like railway obstacle detection, the training dataset is limited and not all possible obstacle classes are known beforehand. For such safety-critical applications, this situation is problematic and could limit the performance of obstacle detection in autonomous trains. In this paper, we propose an exploratory study using unsupervised models based on a large set of generated convolutional autoencoder models to detect obstacles on railway's track level. The study was conducted based on three components: loss functions, activations and optimizers. Existing works rely on fixing thresholds to judge the performance of the model. We propose instead a methodology based on Multi-Criteria Decision Making (MCDM) to evaluate the performance of all models. Furthermore, we introduce the notion of gap-score to evaluate each model by calculating the average difference between the reconstruction score on images with and without obstacles. The aim is to find models maximizing the average of gap-scores and rank them according to their performances. Experimental results show that the evaluated models can provide up to 68% average gap-score.

Local and Global Information in Obstacle Detection on Railway Tracks

“…Gasparini et al [42] combine unsupervised image reconstruction with supervised detection of anomalies for nighttime railway inspection but are thus limited to thermal cameras. Boussik et al [14] perform a grid search over AE structures with different optimizers, activations, and loss functions and evaluate them on a custom test dataset with artificially inserted obstacles and one real-world scenario. In our experiments, Fig.…”

“…Beyond the ability to detect generic obstacles, another advantage of our approach is that the training does not require real or artificially created examples of obstacles on railways. While visual anomaly detection has been extensively studied in the context of industrial inspection [9]- [11] and autonomous driving [12], [13], Boussik et al [14] are the only ones addressing the problem of anomaly detection for railway environments. They proposed training a series of Auto-Encoders (AEs) and using their reconstruction errors as a metric for anomalies.…”

“…We also include random (non-railway) images as negative examples in the training procedure. Inspired by Boussik et al [14], we also study the incorporation of limited global information through obstacle-free images hallucinated by a neural network. Our models are trained on an obstacle-free training set consisting of a subset of RailSem19 [15] and evaluated on an obstacle-enhanced version of it (see Figure 2), following a procedure proposed by Blum et al [12].…”

A Risk-Based Decision-Making Process for Autonomous Trains Using POMDP: Case of the Anti-Collision Function

“…In response to these challenges, the contribution of the paper consists in developing a risk-based 7 decision-making process for the anti-collision function of autonomous trains. The proposed process is able to account for the inherent uncertainty associated with the train state and the wide range of operational and environmental conditions, by using Partially Observable Markov Decision Processes (POMDPs).…”

“…https://www.sncf.com/fr 2 https://railenium.eu train. While these projects encompass numerous engineering research challenges, they primarily focus on the exploration of Artificial Intelligence (AI) techniques[3,4,5] in perception[6,7], control and decision-making functions[8]. Additional focus areas include human-machine cooperation[9], societal acceptance of autonomous technologies[10], as well as risk assessment3 and safety demonstration[12,13,14].The French initiative is part of a global movement to advance autonomous transportation systems.…”