Rudolf Mester

Statistical model-based change detection in moving video

Abstract A major issue with change detection in video sequences is to guarantee robust detection results in the presence of noise. In this contribution, we first compare different test statistics in this respect. The distributions of these statistics for the null hypothesis are given, so that significance tests can be carried out. An objective comparison between the different statistics can thus be based on identical false alarm rates. However, it will also be pointed out that the global thresholding methods resulting from the significance approach exhibit certain weaknesses. Their shortcomings can be overcome by the Markov random field based refining method derived in the second part of this paper. This method serves three purposes: it accurately locates boundaries between changed and unchanged areas, it brings to bear a regularizing effect on these boundaries in order to smooth them, and it eliminates small regions if the original data permits this.

Illumination-Invariant Change Detection Using a Statistical Colinearity Criterion

This paper describes a newalgorithm for illumination-invariant change detection that combines a simple multiplicative illumination model with decision theoretic approaches to change detection. The core of our algorithm is a new statistical test for linear dependence (colinearity) of vectors observed in noise. This criterion can be employed for a significance test, but a considerable improvement of reliability for real-world image sequences is achieved if it is integrated into a Bayesian framework that exploits spatio-temporal contiguity and prior knowledge about shape and size of typical change detection masks. In the latter approach, an MRF-based prior model for the sought change masks can be applied successfully. With this approach, spurious spot-like decision errors can be almost fully eliminated.

Illumination-robust dense optical flow using census signatures

Vision-based motion perception builds primarily on the concept of optical flow. Modern optical flow approaches suffer from several shortcomings, especially in real, non-ideal scenarios such as traffic scenes. Non-constant illumination conditions in consecutive frames of the input image sequence are among these shortcomings. We propose and evaluate the application of intrinsically illumination-invariant census transforms within a dense state-of-the-art variational optical flow computation scheme. Our technique improves robustness against illumination changes, caused either by altering physical illumination or camera parameter adjustments. Since census signatures can be implemented quite efficiently, the resulting optical flow fields can be computed in real-time.

On texture analysis: local energy transforms versus quadrature filters

Abstract The well-known method proposed by Laws for texture analysis first subjects the texture to a filter bank, followed by the computation of energy measures, e.g. through local variance estimation. As shown by Unser in 1986, the filter bank application is equivalent to a linear transformation of the grey values of neighbouring pixels. In this contribution, we derive a further linear relationship between the local variances of the filter outputs and the autocorrelation function of the texture process. Furthermore, we examine how the filter bank approach is related to another method based on multifiltering, namely that one using quadrature filter pairs, by means of which the amplitude envelopes of the filtered texture signal can be obtained. It is shown that the texture energy method can be understood as the equivalent of an envelope detector receiver commonly used in AM communication techniques. Feature images provided by the texture energy method are compared with their counterparts resulting from the quadrature filter approach, and criteria helping to decide when to use which one of the methods are given.

Robust stereo visual odometry from monocular techniques

Visual odometry is one of the most active topics in computer vision. The automotive industry is particularly interested in this field due to the appeal of achieving a high degree of accuracy with inexpensive sensors such as cameras. The best results on this task are currently achieved by systems based on a calibrated stereo camera rig, whereas monocular systems are generally lagging behind in terms of performance. We hypothesise that this is due to stereo visual odometry being an inherently easier problem, rather than than due to higher quality of the state of the art stereo based algorithms. Under this hypothesis, techniques developed for monocular visual odometry systems would be, in general, more refined and robust since they have to deal with an intrinsically more difficult problem. In this work we present a novel stereo visual odometry system for automotive applications based on advanced monocular techniques. We show that the generalization of these techniques to the stereo case result in a significant improvement of the robustness and accuracy of stereo based visual odometry. We support our claims by the system results on the well known KITTI benchmark, achieving the top rank for visual only systems*.

Bayesian illumination-invariant motion detection

We describe an algorithm for change detection which is insensitive to both slow and fast temporal variations of scene illumination. Our algorithm is based on statistical decision theory by using a Bayesian approach. The goal is to detect only temporal changes which are induced by true scene changes, like motion, but not changes due to varying illumination or noise. To this end, our algorithm uses a simple illumination model which is invariant to common camera nonlinearities like gamma-nonlinearity. This is combined with a model for the influence of noise as well as an a priori model for the expected properties of the sought change masks. Key ingredients of the resulting algorithm are a suitable test statistic and an adaptive threshold mechanism. As the algorithm can be applied in a noniterative manner, it is also computationally attractive.

Discriminative Subspace Clustering

We present a novel method for clustering data drawn from a union of arbitrary dimensional subspaces, called Discriminative Subspace Clustering (DiSC). DiSC solves the subspace clustering problem by using a quadratic classifier trained from unlabeled data (clustering by classification). We generate labels by exploiting the locality of points from the same subspace and a basic affinity criterion. A number of classifiers are then diversely trained from different partitions of the data, and their results are combined together in an ensemble, in order to obtain the final clustering result. We have tested our method with 4 challenging datasets and compared against 8 state-of-the-art methods from literature. Our results show that DiSC is a very strong performer in both accuracy and robustness, and also of low computational complexity.

A Maximum Likelihood Estimator for Choosing the Regularization Parameters in Global Optical Flow Methods

Global optical flow estimation methods based on variational calculus contain a regularization parameter which controls the tradeoff between the different constraints on the optical flow field. The counterpart to the regularization parameter are the hyper-parameters in the Bayesian framework. These hyper-parameters have distinct physical meanings and thus can be inferred from the observable data. We derive a combined marginal maximum likelihood/maximum a posteriori (MML/MAP) estimator for simultaneously estimating hyper-parameters and optical flow for all differential variational approaches directly from the observed signal without any prior knowledge of the optical flow. Experiments demonstrate the performance of this optimization technique and show that the choice of the regularization parameter is an essential key-point in order to obtain precise motion estimation.

Detection and description of moving objects by stochastic modelling and analysis of complex scenes

Abstract This paper presents a new technique for the detection and description of moving objects in natural scenes which is based on a statistical multi-feature analysis of video sequences. In most conventional schemes for the detection of moving objects, temporal differences of subsequent images from a video sequence are evaluated by so-called change detection algorithms. These methods are based on the assumption that significant temporal changes of an image signal are caused by moving objects in the scene. However, as temporal changes of an image signal can as well be caused by many other sources (camera noise, varying illumination, small camera motion), such systems are afflicted with the dilemma of either causing many false alarms or failing to detect relevant events. To cope with this problem, the additional features of texture and motion beyond temporal signal differences are extracted and evaluated in the new algorithm. The adaptation of this method to normal fluctuations of the observed scene is performed by a time-recursive space-variant estimation of the temporal probability distributions of the different features (signal difference, texture and motion). Feature data which differ significantly from the estimated distributions are interpreted to be caused by moving objects.

Unbiased Errors-In-Variables Estimation Using Generalized Eigensystem Analysis

Recent research provided several new and fast approaches for the class of parameter estimation problems that are common in computer vision. Incorporation of complex noise model (mostly in form of covariance matrices) into errors-in-variables or total least squares models led to a considerable improvement of existing algorithms.