Obstacle detection is a critical and necessary task for the proper functioning of driverless vehicles. Autonomous driving is a very active field of research, particularly in the automobile industry. However, very few works have been realized for railways obstacle detection especially when it comes to machine learning-based methods, this is often the case because the data is not readily available or the classes for the obstacles are not known beforehand. In this paper, we propose a framework to detect obstacles for autonomous trains using deep learning methods trained on frontal ego-perspective acquisitions of the train. Using normal railways scene images, we aim to exploit deep learning methods to detect obstacles as anomalies on the main rail track. In our approach, we do not assume any prior knowledge of the obstacles' classes by using unsupervised learning. Our contribution is threefold: First, we propose an algorithm to detect and localize all rails in input images. Second, we propose a pipeline based on clustering on the previously detected rails to extract, in real-time, the region of interest i.e the main rail track. Third, in order to detect obstacles, we propose a convolutional auto-encoder based method to process the aforementioned regions of interest. Our framework was tested and validated on railway scenes consisting of a mixture of inlaid obstacles and real-world scenarios containing real obstacles.