Browsing by Author "Brink, Daniek"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- ItemUsing probabilistic graphical models to detect dynamic objects for mobile robots(Stellenbosch : Stellenbosch University, 2016-12) Brink, Daniek; Van Daalen, C. E.; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.ENGLISH ABSTRACT: An autonomous mobile robot must be able to identify moving objects in its environment, continually as it is operating, for accurate environment mapping and collision-free navigation. This is not an easy task, since most of what might be observed will appear to be moving due to the robot’s own motion. The task is further complicated by the inherent uncertainty in the pose estimates and environment measurements captured by the robot. In this work we focus on features in the environment whose 3D locations are measured over time, such as triangulated stereo image features. Our aim is to separate dynamic features from static ones and also to group the dynamic ones into separate objects. Existing approaches generally assume that the exact pose of the robot is known at every time step or, in order to estimate the robot pose, they assume that the environment is predominantly stationary. We avoid these assumptions through thoughtful consideration for the uncertainties involved. In order to model the uncertainties, as well as the statistical dependencies between observations and latent variables, we present a novel application of probabilisitic graphical models (PGMs) for dynamic object detection. Our PGM can be divided into two interacting components. The first relates to motion segmentation, in which all observed features are classified as static or dynamic, and the second relates to object segmentation, in which dynamic features on the same objects are clustered together. We also take care to accommodate for semi-static objects, which are objects that can be both stationary and dynamic during the observation period. Our design choices lead to a PGM containing both discrete and continuous variables. Tractable inference in such a hybrid model can be challenging, and we pay particular attention to this issue. It turns out that messages sent from continuous to discrete variables can be pre-computed, before loopy belief propagation is performed over the discrete variables. Experiments on the KITTI benchmark datasets indicate that our PGM approach performs well, and it often outperforms a state-of-the-art feature-based algorithm. We find that motion segmentation accuracy tends to improve as more observations of the same features become available, and that our method has the ability to handle semi-static objects successfully. The ability of our PGM to segment different objects is also seen to perform superior.