Automatic methods are needed to efficiently process the large point clouds collected using a mobile laser scanning (MLS) system for surveying applications. Machine-learning-based object recognition from MLS point clouds in a road and street environment was studied in order to create maps from the road environment infrastructure. The developed automatic processing workflow included the following phases: the removal of the ground and buildings, segmentation, segment classification, and object location estimation. Several novel geometry-based features, which were previously applied in autonomous driving and general point cloud processing, were applied for the segment classification of MLS point clouds. The features were divided into three sets, i.e., local descriptor histograms (LDHs), spin images, and general shape and point distribution features, respectively. These were used in the classification of the following roadside objects: trees, lamp posts, traffic signs, cars, pedestrians, and hoardings. The accuracy of the object recognition workflow was evaluated using a data set that contained more than 400 objects. LDHs and spin images were applied for the first time for machine-learning-based object classification in MLS point clouds in the surveying applications of the road and street environment. The use of these features improved the classification accuracy by 9.6% (resulting in 87.9% accuracy) compared with the accuracy obtained using 17 general shape and point distribution features that represent the current state of the art in the field of MLS; therefore, significant improvement in the classification accuracy was achieved. Connected component segmentation and ground extraction were the cause of most of the errors and should be thus improved in the future.

We consider the problem of locating instances of a known object in a novel scene by matching the fiducial features of the object. The appearance of the features and the shape of the object are modeled separately and combined in a Bayesian framework. In this paper, we present a novel matching scheme based on Sequential Monte Carlo, in which the features are matched sequentially, utilizing the information about the locations of previously matched features to constrain the task. The particle representation of hypotheses about the object position allow matching in multimodal and cluttered environments, where batch algorithms may have convergence difficulties. The proposed method requires no initialization or predetermined matching order, as the sequence can be started from any feature. We also utilize a Bayesian model to deal with features that are not detected due to occlusions or abnormal appearance. In our experiments, the proposed matching system shows promising results, with performance equal to batch approaches when the target distribution is unimodal, while surpassing traditional methods under multimodal conditions. Using the occlusion model, the object can be localized from only a few visible features, with the nonvisible parts predicted from the conditional prior model.

We consider the problem of locating instances of a known object in a novel scene by matching the fiducial features of the object. Our approach to the problem consists of two parts: a model for the appearance of the features and a model for the shape of the object. We then bind these parts together in a Bayesian framework and match the features sequentially, using the information about the locations of previously matched features. Into this matching system we add a Bayesian model for dealing with features that are not detected due to occlusion or abnormal appearance. Our system yields promising results, losing little matching accuracy even for heavily occluded objects. 1

We present methods to obtain computationally efficient proposal distributions for Bayesian reversible jump Markov chain Monte Carlo (RJMCMC) based image segmentation. The slow convergence of MCMC methods often makes them poorly suited for practical image processing applications. We show how carefully crafted proposal distributions along with certain approximations can decrease the computational cost of MCMC image segmentation to a level that is comparable with some traditional algorithms. We also discuss the interpretation of the resulting distribution of different segmentations and present experimental results.

This article presents a classical type of solution to the time series prediction competition, the CATS benchmark, which is organized as a special session of the IJCNN 2004 conference. The solution is based on sequential application of the Kalman smoother, which is a classical statistical tool for estimation and prediction of time series. The Kalman smoother belongs to the class of linear methods, because the underlying filtering model is linear and the distributions are assumed as Gaussian. Since the time series model of the Kalman smoother assumes that the densities of noise terms are known, these are determined by cross-validation.

This article presents a classical type of solution to the time series prediction competition, the CATS benchmark, which is organized as a special session of the IJCNN 2004 conference. The solution is based on sequential application of the Kalman smoother, which is a classical statistical tool for estimation and prediction of time series. The Kalman smoother belongs to the class of linear methods, because the underlying filtering model is linear and the distributions are assumed as Gaussian. Since the time series model of the Kalman smoother assumes that the densities of noise terms are known, these are determined by cross-validation.

In this work, we discuss practical methods for the assessment, comparison, and selection of complex hierarchical Bayesian models. A natural way to assess the goodness of the model is to estimate its future predictive capability by estimating expected utilities. Instead of just making a point estimate, it is important to obtain the distribution of the expected utility estimate because it describes the uncertainty in the estimate. The distributions of the expected utility estimates can also be used to compare models, for example, by computing the probability of one model having a better expected utility than some other model. We propose an approach using cross-validation predictive densities to obtain expected utility estimates and Bayesian bootstrap to obtain samples from their distributions. We also discuss the probabilistic assumptions made and properties of two practical cross-validation methods, importance sampling and k-fold cross-validation. As illustrative examples, we use multilayer perceptron neural networks and gaussian processes with Markov chain Monte Carlo sampling in one toy problem and two challenging real-world problems.

In this paper we show that rotational invariance can be improved in a neural network based electrical impedance tomography (EIT) reconstruction approach by a suitably chosen permutation of the input data. The input space is partitioned to nonoverlapping sectors, and the input signal is permuted so that it lies in one sector independent of the original rotation angle. We demonstrate the advantages of the method with computer simulations. The proposed approach yields better results in the inverse problem, and allows use of smaller networks with fewer training samples.

Bayesian MLP neural networks are a flexible tool in complex nonlinear problems. The approach is complicated by need to evaluate integrals over high-dimensional probability distributions. The integrals are generally approximated with Markov chain Monte Carlo (MCMC) methods. There are several practical issues which arise when implementing MCMC. This article discusses the choice of starting values and the number of chains in Bayesian MLP models. We propose a new method for choosing the starting values based on early stopping and we demonstrate the benefits of using several independent chains.

We discuss the importance of finding the correct model complexity and regularization level, in the self-organizing map (SOM) algorithm. The complexity of the SOM is determined mainly by the width of the final neighborhood, which is usually chosen ad hoc or set to zero for optimal quantization error. However, if the SOM is used for visualizing the joint probability distribution of the data, then care must be taken not to overfit the model to the data sample, similarly as with any statistical model. We propose a heuristic criterion for model selection in SOM, and demonstrate by simulations that the criterion can be used for selecting the neighborhood that suppresses overfitting.

Training with hints is a powerful method for incorporating almost any type of prior knowledge into neural network models. In this paper we demonstrate how the hints can be constructed from numerical approximation of the regularization cost function, and discuss the problem of selecting the hint samples. We give a simple algorithm for placing the hint samples in such regions in the input space where the hint error is large, and for selecting the minimum sufficient set of hint samples by removing the correlated samples.

In this contribution we present a method for solving the inverse problem in electric impedance tomography with neural networks. The problem of reconstructing the conductivity distribution inside an object from potential measurements on the surface is known to be ill-posed requiring efficient regularization techniques. We demonstrate that a statistical inverse solution, where the mean of the inverse mapping is approximated with a neural network gives promising results. We study the effect of input and output data representation by simulations and conclude that projection to principal axis is feasible data transformation. Also we demonstrate that Bayesian neural networks, which aim to average over all network models weighted by the model's posterior probability provide the best reconstruction results. With the presented approach estimation of some target variables, such as the void fraction (the ratio of gas and liquid), may be applicable directly without the actual image reconstruction. We also demonstrate that the solutions are very robust against noise in inputs.

In a multivariate regression problem it is often assumed that residuals of outputs are independent of each other. In many applications a more realistic model would allow dependencies between the outputs. In this paper we show how a Bayesian treatment using the Markov chain Monte Carlo method can allow for a full covariance matrix with multilayer perceptron neural network.

A method for constructing classification features with unsupervised learning is presented. The method is based on clustering of the high dimensional measurements into a small number of features with self-organizing maps. The histograms of the self-organized features are classified with a multilayer perceptron network, that can pick up the relevant features and feature combinations from the histograms. The method is applied in two industrial problems, color image recognition for selection of optimal reproduction parameters in the printing press, and defect classification in wood surfaces. In both applications, the results were evaluated by the domain experts to be sufficient with respect to the application requirements. In the color image recognition the results were compared to the manually selected parameter settings, and in 12% of the test images there were distinquishable differences, with very few clear failures. Performance of the wood defect classification system was evaluated with a set of 400 knot images with 7 classes from spruce boards. The recognition rate was about 85% with only gray level images, giving about 90% accuracy for the final board grading, to be compared to 75–80% accuracy that can be maintained by a human inspector. As the methods are based on rather generic basic features and learning of the final features and classes from the samples, they are easily adaptable to different tasks, such as defect inspection in lumber or veneer, with different tree species or different cutting processes.

A generic, modular, neural network-based feature extraction and pattern classification system is proposed for finding essentially two-dimensional objects or object parts from digital images in a distortion tolerant manner, The distortion tolerance is built up gradually by successive blocks in a pipeline architecture. The system consists of only feedforward neural networks, allowing efficient parallel implementation. The most time and data-consuming stage, learning the relevant features, is wholly unsupervised and can be made off-line. The consequent supervised stage where the object classes are learned is simple and fast. The feature extraction is based on distortion tolerant Gabor transformations, followed by minimum distortion clustering by multilayer self-organizing maps. Due to the unsupervised learning strategy, there is no need for preclassified training samples or other explicit selection for training patterns during the training, which allows a large amount of training material to be used at the early stages, A supervised, one-layer subspace network classifier on top of the feature extractor is used for object labeling. The system has been trained with natural images giving the relevant features, and human faces and their parts have been used as the object classes for testing. The current experiments indicate that the feature space has sufficient resolution power for a moderate number of classes with rather strong distortions.

A local spatial convolution filter is suggested whose design is based on approximating a global optimal restoration filter, the Parametric Projection Filter (PPF). The PPF tries to construct a restored signal or image that is optimally close to each individual original signal, with the noise level as low as possible. The local filter provides a restoration that is as close as possible to the PPF restoration. The design is accomplished by least squares techique in Fourier domain. Once the PPF is available, the main task is to solve a set of linear normal equations whose coefficient size depends on the size of the local filter. It is shown by examples using both one- and two-dimensional signals that the restorations are satisfactory even when the filter size is considerably smaller than the width of the degradation.