Visual recognition of bridges by using stereo cameras on trains

Recognition of either patterns or objects in mobile systems continues to be in the focus of intensive research, with many applications being enhanced by integrating environment related information. This paper presents a practical technique for detecting and recognizing bridges from a train using a stereo camera which provides depth and grayscale images. The algorithm has been applied to a train system, where object detection combined with a given map of an area is used to improve localization. The approach is based on the detection of primitive features including edges and corners in the depth image. The pairwise spatial relations between the features are then modeled by a graph, so the classification and detection can be performed by a probabilistic Markov Random Field framework. The algorithm has been tested on the real-life datasets of the Rail Collision Avoidance System (RCAS) project. The presented results prove the applicability of the framework for detection of objects by exploiting geometrical appearance constraints.