Traversability analysis using terrain mapping and online-trained Terrain type classifier

Abstract

Abstract-Path estimation is a big challenge for autonomous vehicle navigation, especially in unknown, dynamic environments, when road characteristics change often. 3D terrain information (e.g. stereo cameras) can provide useful hints about the traversability cost of certain regions. However, when the terrain tends to be flat and uniform, it is difficult to identify a better path using 3D map solely. In this scenario the use of a priori knowledge on the expected road's visual characteristics can support detection, but it has the drawback of being not robust to environmental changes. This paper presents a path detection method that mixes together 3D mapping and visual classification, trying to learn, in real time, the actual road characteristics. An on-line learning of visual characteristics is implemented to feedback a terrain classifier, so that the road characteristics are updated as the vehicle moves. The feedback data are taken from a 3D traversability cost map, which provides some hints on traversable and non-traversable regions. After several re-training cycles the algorithm converges on a better separation of the path and non-path regions. The fusion of both 3D traversability cost and visual characteristics of the terrain yields a better estimation when compared with either of these methods solely.

Self-Supervised Learning for Autonomous Vehicles Perception: A Conciliation Between Analytical and Learning Methods

“…2. Pre-trained SSL urban road systems: Some other online SSL techniques deal with this issue by exploiting a classifer pre-trained offline on hand-labeled data [50], [51].…”

“…A hybrid path segmentation technique is proposed in [51]. It combines a 3D traversability cost map obtained by stereo-vision, and an SVM classifier pre-trained offline over a human annotated dataset.…”

“…To sum up these road segmentation SSL approaches, we can notice that while the sensor data and the analytical and learning models are different, the online process remains essentially the same. The first stage always consists in generating automatic labels by using additional temporal [49], sensor [13], [19], or prior knowledge information [50], [51]. Then, a second stage trains or updates online a classifier, such that it can be used to provide a long-range or refine road segmentation.…”

Learning-Based Methods of Perception and Navigation for Ground Vehicles in Unstructured Environments: A Review

“…An example of hybrid approach is given by Roncancio et al [ 62 ] who propose a path detection method. An on-line learning of visual features (or texture descriptors) is implemented to continually update an SVM-based terrain classifier, in order to enhance the path detection as the vehicle moves.…”

Watch Your Step! Terrain Traversability for Robot Control

“…Among the number of methods and models for terrain analysis, there are at least two large categories, (i) classification-based methods and (ii) cost-assessment methods. In the former, it is possible to count all the approaches that consider a binary distinction of the terrain as two classes, traversable or non-traversable; to cite an example, in [7], the authors use an on-line trained classifier to distinguish traversable and non-traversable regions. Widely spread in research, occupancy maps also fall in this category as they use the elevation of surrounding objects to construct a map of occupied regions on the base of sensor measurements [8].…”