川本 一彦

© Springer-Verlag Berlin Heidelberg 2014. We develop an algorithm for the computation of a locally affine optical flow field as an extension of the Lucas-Kanade (LK) method. The classical LK method solves a system of linear equations assuming that the flow field is locally constant. Our method solves a collection of systems of linear equations assuming that the flow field is locally affine. Since our method combines the minimisation of the total variation and the decomposition of the region, the method is a local version of the l² <inf>2</inf>-l<inf>1</inf> optical flow computation. Since the linearly diverging vector field from a point is locally affine, our method is suitable for optical flow computation for diverging image sequences such as front-view sequences observed by car-mounted cameras.

The purpose of this paper is twofold. First, we introduce fast global image registration using random projection. By generating many transformed images as entries in a dictionary from a reference image, nearest-neighbour-search (NNS)-based image registration computes the transformation that establishes the best match among the generated transformations. For the reduction in the computational cost for NNS without a significant loss of accuracy, we use random projection. Furthermore, for the reduction in the computational complexity of random projection, we use the spectrum-spreading technique and circular convolution. Second, for the reduction in the space complexity of the dictionary, we introduce an interpolation technique into the dictionary using the linear subspace method and a local linear property of the pattern space.

The purpose of this paper is three-fold. First, we develop an algorith for the computation a locally affine optical flow field from multichannel images as an extension of the Lucus-Kanade (LK) method. The classical LK method solves a system of linear equations assuming that the flow field is locally constant. Our method solves a collection of systems of linear equations assuming the flow field is locally affine. For autonomous navigation in a real environment, the adaptation of the motion and image analysis algorithm to illumination changes is a fundamental problem, because illumination changes in an image sequence yield counterfeit obstacles. Second, we evaluate the colour channel selection of colour optical flow computation. By selecting an appropriate colour channel, it is possible to avoid these counterfeit obstacle regions in the snapshot image in front of a vehicle. Finally, we introduce an evaluation criterion for the computed optical flow field without ground truth.

We present an algorithm for scale-space clustering of point cloud on the sphere using the methodology for the estimation of the density distribution of the points in the linear scale space. Our algorithm regards the union of observed point sets as an image defined by the delta functions located at the positions of the points on the sphere. A blurred version of this image has a deterministic structure which qualitatively represents the density distribution of the points in a point cloud on a manifold.

This paper consider the problem of handling appearance variability in visual tracking and proposes an appearance generative model for visual vehicle tracking. The generative model is used to adaptively generate and update the appearance templates during visual tracking. The appearance templates are efficiently represented in a low dimensional eigen subspace learned from pre-acquired templates and are parameterized by two pose parameters of a target object. The adaptive template updating is made by particle filtering in which the particles represents the appearance templates. In experiments with real image sequences, we show the effectiveness of the proposed method.

This paper is concerned with the problem of estimating both continuous and discrete hidden states of stochastic dynamic systems, called hybrid state space models. Of several hybrid state space models, this paper focuses on switching observation models. In this model, the discrete states are used for indicating one of several observation models, and the continuous and discrete states independently evolve with time but the observations is generated depending on both the states. The contribution of this paper is to propose a Rao-Blackwellised particle filter for this model in conjunction with the grid based filter. Normally, the Kalman filter has been most widely used for Rao-Blackwellisation of particle filters, but in the proposed algorithm the grid-based filter is used to analytically estimate the posterior distribution on the discrete state. In experiments a nonlinear and non-Gaussian benchmark model is used to evaluate the performance of the proposed algorithm. The experimental result with Monte Carlo simulations shows that the proposed algorithm outperforms the basic particle filter, especially when the number of the particles used for estimation is small. © 2012 ISIF (Intl Society of Information Fusi).

The most important part of an information system that assists human activities is a natural interface with human beings. Gaze information strongly reflects the human interest or their attention, and thus, a gaze-based interface is promising for future usage. In particular, if we can smoothly guide the user's visual attention toward a target without interrupting their current visual attention, the usefulness of the gaze-based interface will be highly enhanced. To realize such an interface, this paper proposes a method for editing an image, when given a region in the image, to synthesize the image in which the region is most salient. Our method first computes a saliency map of a given image and then iteratively adjusts the intensity and color until the saliency inside the region becomes the highest for the entire image. Experimental results confirm that our image editing method naturally draws the human visual attention toward our specified region. © 2011 ACM.

An evaluation strategy for visual key image retrieval systems is proposed in order to show the design criteria of a querying interface on mobile devices. Indexes (lists of visual keys) generated by different parameters are analyzed, and indexes derived by different visual keys are validated using ArtExplosion 600,000, which contains about 300 semantic categories and over 100,000 natural photos. The result suggests that access to a collection with a visual key can provide one to three relevant images in rank 10 when the number of visual keys is 60, which is the lower limit. In portable devices, which can display 16 visual keys per page, users can at least access a required image by browsing only 4 pages with 60 visual keys, and can use the image for related subsequent queries by using the other image retrieval functions.

This paper concerns the problem of estimating both continuous and discrete hidden states of stochastic dynamic systems, which model is called hybrid state space model, and especially focuses on a class of the model with switching observation models. In this model, discrete states are used for indicating one of the given observation models. This class of model naturally appears in applications such as multiple sensor tracking, data association for multiple targets tracking, and outlier detection of time series signals. The contribution of this paper is to propose a particle filter for this model in conjunction with hidden Markov model (HMM) filter. The HMM filter is used for estimating discrete states and provides the optimal solution under given continues states from the Bayesian point of view. Experimental results show the effectiveness of the proposed method.

静止したシーンを移動しながら撮影した動画像から内部パラメータ既知のカメラ運動を逐次的に推定するための粒子フィルタアルゴリズムを提案する.提案手法は,回転3自由度を単位回転軸の球面座標とその周りの回転角度及び単位並進2自由度を球面座標でパラメータ化し,この運動パラメータ空間で直接粒子(サンプル)を生成し運動を推定する.従来法の多くは,2画像間の特徴点対応から基本行列を推定し回転と並進に分解する逆問題を解こうとするが,提案手法の手順はその逆になり,生成的なアプローチになる.この手順により,最適化問題を解くことなく,推定した運動パラメータから構城した基本行列は自動的に分解可能条件を満たしエピ極線は1点で交わる.更に,並進運動がないなどの退化運動に対処するために運動モデル切換の仕組みなどいくつかの工夫を導入する.シミュレーション動画像と実動画像を用いた実験により,提案手法の有効性を検証する.

We propose a method for recursively estimating the parameters of a numerical simulation model for pedestrian motion using an image sequence. We construct the model with so-called social forces, which have been successfully used in computer simulations for pedestrian motion analysis. The contribution of this paper is to combine the numerical simulation model and observations captured from image sequences. To this end, we introduce the framework of data assimilation, which is originally developed in geosciences such as weather forecasting and hydrology for refining numerical simulation models using observations available in the real world. In addition we use a particle filter for the recursive Bayesian estimation In experiments with real videos [9] we show a case study of pedestrian motion analysis. © 2010 TSI Press.

This paper treats the problem of tracking objects of interest in video sequences, and proposes an appearance-based algorithm for visual tracking with a particle filter. The algorithm can find geometric transformation parameters, such as affine, similar, and Euclidean transformations, which describe the appearance change of a given object between two successive images. Since such geometric transformations contains nonlinear functions such as trigonometric functions and multiplication, linear and Gaussian state space algorithms, namely Kalman filters, can not directly address the dynamics with the transformations. Thus the algorithm is implemented with a particle filter, which is capable of dealing with nonlinear and non-Gaussian state space models using Monte Carlo approximation. The algorithm based on state space models explicitly deal with the dynamics of targets in video sequences, while widely used deforomable template matching algorithms depends on heuristic search. Experimantal results with a real video sequence of a traffic scene are shown to evaluate the performance. The modeling technology gives a basis for traffic analysis such as congestion prediction and accident monitoring in Intelligent Transport Systems.

A traffic accident prediction method using a priori knowledge based on accident data is proposed for safe driving support. Implementation is achieved by an algorithm using particle filtering and fuzzy inference to estimate accident risk factors. With this method, the distance between the host vehicle and a vehicle ahead and their relative velocity and relative acceleration are obtained from the results of particle filtering of driving data and are used as attributes to build the relative driving state space. The attributes are evaluated as likelihoods and then consolidated as a risk level using fuzzy inference. Experimental validation was done using videos of general driving situations obtained with an on-vehicle CCD camera and one simulated accident situation created based on the video data. The results show that high risk levels were calculated with the proposed method in the early stages of the accident situations.

A random scanning algorithm for tracking curves in image sequences is proposed. First it verifies the presence of a curve by votes, whose values indicate the number of the points on the curve, thus being able to robustly detect curves against outlier and occlusion. Furthermore, a procedure for preventing redundant verification is introduced by determining equivalence curves in the digital space to reduce the time complexity. Second it propagates the distribution which represents the presence of the curve to the successive image of a given sequence. This temporal propagation enables to focus on the potential region where the curves detected at time t-1 are likely to appear at time t. As a result, the time complexity does not depend on the dimension of the curve to be detected. To evaluate the performance, we use three noisy image sequences, consisting of 90 frames with 320 x 240 pixels. The results show that the algorithm successfully tracks the target even in noisy or cluttered binary images. © 2008 Taylor & Francis Group, LLC.

We propose a statistical motion model for sequential Bayesian tracking and show an adaptive particle filter algorithm for the motion model. It predicts the current state with the help of optical flows, i.e., it explores the state space with information based on the current and previous images of an image sequence. In addition, we introduce a robust method for state estimation and an automatic method for adjusting the variance of the motion model, which parameter is manually determined in most particle filters. In experiments with a real image sequence, we compare the proposed motion model with a random walk model, which is a widely used model for tracking, and show the proposed model outperform the random walk model.

線形推定に基づく逐次重点サンプリング型のパーティクルフィルタを提案する.従来広く用いられている事前モデル型サンプリング法では,急激な状態変化が発生したときに予測が外れ,事後分布の近似精度が劣化してしまう.提案手法は,線形推定による大まかな予測に基づいて重点サンプリングするため追従性に優れている.まず,線形最小二乗法に基づく重点関数の設計法を提案し,優決定の連立方程式を毎時刻解く必要がない方法を示す.次に,誤追跡による外れ値を考慮して線形ロバスト推定に基づく方法へ拡張し,数値探索に頼らない方法を示す.最後に,大きな状態変化の検出により事前モデルと提案する重点関数を切り換える方法を提案する.これにより,滑らかな推定と急な状態変化の検出の両方を扱うことができる.急な運動変化を含む実画像に対して従来手法と比較し,増加する計算時間はたかだか1ms/frameでありながら,高精度かつ頑健な推定を達成していること示す.

This paper proposes an algorithm for detecting objects representing potential hazards to drivers based on the combination of local information derived from optical flows and global information obtained from the host vehicle's status. The algorithm uses artificial neural networks to infer the degree of danger posed by moving objects in dynamic images taken with a vehicle-mounted camera. This approach allows more flexible adaptation of the algorithm to many drivers with dissimilar characteristics. Experiments were conducted with both miniature vehicles in a virtual environment and real vehicles in a real driving situation using video images of multiple moving objects. The results show that the algorithm can infer hazardous situations similar to the judgments made by human drivers. The proposed algorithm provides the foundation for constructing a practical driving assistance system.

A new concept of drive recording system based on estimation of human intention by particle filters has been proposed. It has an omni-directional camera at front ceiling in car cabin to record outside cars and inside driver simultaneously. Outside cars are detected with simple image processing and tracked by finite random set (FRS) state space approach, while driver's head direction is detected by eigen space method and tracked by elaborated particle filtering technique. Noticed cars noticed of the driver are determined by matching driver's head direction with the outside cars. The system estimates driver's prediction on the outside cars' motion based on the detection result of noticed cars. Hiyari-hatto incident will be detected when large discrepancy between the estimated prediction and the actual situation occurred. Then the system records image sequences around the incident time as well as other signals obtained from sensors of the car. © 2007 IEEE.

We propose a statistical motion model for sequential Bayesian tracking, called the optical flow-driven motion model, and show an adaptive particle filter algorithm with the motion model. It predicts the current state with the help of optical flows, i.e., it explores the state space with information based on the current and previous images of an image sequence. In addition, we introduce an automatic method for adjusting the variance of the motion model, which parameter is manually determined in most particle filters. In experiments with synthetic and real image sequences, we compare the proposed motion model with a random walk model, which is a widely used model for tracking, and show the proposed model outperform the random walk model in terms of accuracy even though their execution times are almost the same.

A non-invasive technique for local gas holdup measurement of a bubble column using Self-Organized Network inspired by Immune Algorithm (SONIA) neural network and ultrasonic method is investigated. The energy attenuation and the transmission time difference of ultrasound are used as measurement parameters to obtain the local gas holdup in an air-water dispersion system using SONIA neural network reconstruction. Bubble size distributions in the bubble column are obtained by using a photographic method. The experimental results and simulations on three different mean bubble size condition show that the errors of SONIA neural network method is 1/9 times lower than those of the conventional back-propagation neural network. The results show a good agreement with measured data. © 2005 Elsevier B.V. All rights reserved.

A fuzzy-similarity-based self-organized network inspired by immune algorithm (F-SONIA) is proposed in order to develop an artificial odor discrimination system for three-mixture-fragrance recognition. It can deal with an uncertainty in frequency measurements, which is inherent in odor acquisition devices, by employing a fuzzy similarity. Mathematical analysis shows that the use of the fuzzy similarity results on a higher dissimilarity between fragrance classes, therefore, the recognition accuracy is improved and the learning time is reduced. Experiments show that F-SONIA improves recognition accuracy of SONIA by 3% - 9% and the previously developed artificial odor discrimination system by 14% - 25%. In addition, the learning time of F-SONIA is three times faster than that of SONIA. © 2006 IEEE.

Linear estimation based sequential importance sampling methods for particle filters are proposed that can be used to model the rapid change of object motion in a video sequence. First a linear least-squares estimation is used to build a proposal function from observations, and then it is extended to a robust linear estimation. These sampling methods give a framework for tracking objects whose motion cannot be well modeled by a prior model. Finally a switching algorithm between the proposed method and the prior model based sampling method is proposed to achieve a filtering of both smooth and rapid evolution of the state. The ability of the proposed method is illustrated on a real video sequence involving a rapidly moving object.

To improve recognition and generalization capability of back-propagation neural networks (BP-NN), a hidden layer self-organization inspired by immune algorithm called SONIA, is proposed. B cell construction mechanism of immune algorithm inspires a creation of hidden units having local data recognition ability that improves recognition capability. B cell mutation mechanism inspires a creation of hidden units having diverse data representation characteristics that improves generalization capability. Experiments on a sinusoidal benchmark problem show that the approximation error of the proposed network is 1/17 times lower than that of BP-NN. Experiments on real time-temperature-based food quality prediction data shows that the recognition capability is 18% improved comparing to that of BP-NN. The development of the world first time-temperature-based food quality prediction demonstrates the real applicability of the proposed method in the field of food industry. © 2004 Elsevier B.V. All rights reserved.

An algorithm for color restoration under multiple luminance conditions is proposed. It automatically produces correction vectors to restore the color information in the Lz.ast;az.ast;bz.ast; color metric space, using color values of a target object within the well-illuminated region in a given dynamic image. The use of the correction vectors provides better image quality than that obtained by the restoration algorithm using color change vectors. An experiment is done with two real dynamic images, where a walking person in a building is observed, to evaluate the performance of the proposed algorithm in terms of color-difference. The experimental results show that the restored image by the proposed algorithm decreases the color-difference by 30% compared to the restoration algorithm using color change vectors. The proposed algorithm presents the foundation to identify the person captured by a practical security system using a low cost CCD camera. © 2004 Elsevier B.V. All rights reserved.

To deal with a problem of overlapping data in pattern classification, a class majority method in designing hidden units of Fuzzy Local Approximation NN is proposed. Moreover, to improve the output confidence of the networks, Euclidean fuzzy similarity is proposed as hidden unit operator. For each cluster formed by the adaptive clustering of F-SONIA, the number of vectors that belong to the same class is calculated. Therefore the fractions of each class are known. One class should have a single class majority in the cluster. Then, the cluster with no single majority is broken down into two clusters. In experiments, the real-world benchmark datasets, e.g., 2-D vowel, Iris, and thyroid data that have different challenges to the networks in terms of overlapping and size are used to test the networks. Experiments show that the proposed methods improve the classification performance as well as the output confidence of the networks. © 2005 TFSA.