Building large-scale 2.5D maps when spatial correlations are considered can be quite expensive, but there are clear advantages when fusing data. While optimal submapping strategies have been explored previously in covariance-form using Gaussian Process for large-scale mapping, this paper focuses on transferring such concepts into information form. By exploiting the conditional independence property of the Gaussian Markov Random Field (GMRF) models, we propose a submapping approach to build a nearly optimal global 2.5D map. In the proposed approach data is fused by first fitting a GMRF to one sensor dataset; then conditional independent submaps are inferred using this model and updated individually with new data arrives. Finally, the information is propagated from submap to submap to later recover the fully updated map. This is efficiently achieved by exploiting the inherent structure of the GMRF, fusion and propagation all in information form. The key contribution of this paper is the derivation of the algorithm to optimally propagate information through submaps by only updating the common parts between submaps. Our results show the proposed method reduces the computational complexity of the full mapping process while maintaining the accuracy. The performance is evaluated on synthetic data from the Canadian Digital Elevation Data.

2.5D maps are preferable for representing the environment owing to their compactness. When noisy observations from multiple diverse sensors at different resolutions are available, the problem of 2.5D mapping turns to how to compound the information in an effective and efficient manner. This paper proposes a generic probabilistic framework for fusing efficiently multiple sources of sensor data to generate amendable, high-resolution 2.5D maps. The key idea is to exploit the sparse structure of the information matrix. Gaussian Markov Random Fields are employed to learn a prior map, which uses the conditional independence property between spatial location to obtain a representation of the state with a sparse information matrix. This prior map encoded in information form can then be updated with other sources of sensor data in constant time. Later, mean state vector and variances can be also efficiently recovered using sparse matrices techniques. The proposed approach allows accurate estimation of 2.5D maps at arbitrary resolution, while incorporating sensor noise and spatial dependency in a statistically sound way. We apply the proposed framework to pipe wall thickness mapping and fuse data from two diverse sensors that have different resolutions. Experimental results are compared with three other methods, showing that, while greatly reducing computation time, the proposed framework is able to capture in large extend the spatial correlation to generate equivalent results to the computationally expensive optimal fusion method in covariance form with a Gaussian Process prior.

This work exploits modeling spatial correlation in 2.5D data using Gaussian Processes (GPs), and produces constrained sampling realizations on these models to improve certainty in the predictions by means of integrating additional sparse information. Data organized in 2.5D such as elevation and thickness maps has been extensively studied in the fields of robotics and geostatistics. These maps are typically represented as a probabilistic 2D grid that stores an estimated value (height or thickness) for each cell. With the increasing popularity and deployment of robotic devices for infrastructure inspection, 2.5D data becomes a common interpretation of the condition of the target being inspected. Modeling the spatial dependencies and making inferences on new grid locations is a common task that has been addressed using GPs, but inference results on locations which are weakly correlated with the training data are generally not sufficiently informative and distinctly uncertain. The predictive capability of the proposed framework, which is applicable to any 2.5D data, is demonstrated with field inspection data from pipelines. Specifically, sparse and complementary measurements from alternative sensing modalities have been incorporated into the model to predict in more detail local thickness conditions where GP training data is limited. The output of this work aims to probabilistically present variations of the target in the case that both accuracy and reasonable diversity are of significant interest.

Typically 2.5D maps provide a compact and efficient representation of the environment. When sensor data is obtained from multiple sets of noisy measurements at differing resolutions, the problem of compounding this information together to provide an effective and efficient means of mapping is not trivial, particularly as the size of the environment increases. In this paper, we propose a general framework for integrating heterogeneous sensor data to obtain large-scale 2.5D probabilistic maps. Gaussian Processes are used to generate a prior map that learns the spatial correlation between nearby points. Bayesian data fusion is then employed to update these prior maps with new measurements from distinct sensor modalities. In order to deal with large scale data, a novel submapping strategy is introduced to perform the fusion step efficiently in dealing with large covariance matrices. Submaps are first marginalised from the learned correlated prior and then updated based on the property of conditional independence. Most notably, the technique lends itself to generate accurate estimates at arbitrary resolutions and is able to handle varying noise from disparate sensor sources. The framework is applied to pipeline thickness mapping, with experimental results in fusing a high-resolution sensor and a low-resolution sensor showing the ability of the proposed technique to capture spatial correlations to come up with more accurate results when compared with a naïve fusion approach.

We here study the problem of visual attention computation in video of driving environment via the learning from eye movements. We collect a large-scale database of eye movements from 28 subjects on 30 videos of road scenes, which simulate the driving environment. The analysis on this eye movement database reveals that visual attention in driving environment is directed by high-level cognitive factors such as objects. We then present a new high-level representation called Traffic Object Bank (TOB), which is comprised of many individual road object detectors trained comprehensively in semantic space as well as viewpoint space. TOB provides semantically rich object-level features. Finally, we develop a computational model to predict where drivers look via the mapping from TOB-based representation and to gaze data. Experimental results on our traffic scene video benchmark indicate high accordance with human eye movement and show great promise for further applications.

Image segmentation is a critical issue in image understanding and several achievements have been achieved in this area. In this paper, we propose an improved image segmentation algorithm based on mean shift mode detection and Tree-Structured MRF (TS-MRF) model, which we believe is better. Section.2 briefs the mean shift mode detection and the speeded up KNN based mean shift algorithm used in the experiment. Section.3 presents the background knowledge of MRF model and TS-MRF model. On the basis of section.2 and section.3, we apply mean shift algorithm with bandwidth parameter h to calculate the number of clusters n. Then, we set n as the input parameter in MRF based segmentation and get n children nodes. The structure of the tree is formed by conducting above procedures iteratively. The result, shown in Fig.8 through 12 and their analysis demonstrate preliminarily that our novel algorithm is significantly better and still can be improved further.