Daigo Muramatsu

In this paper, we report an implementation of self-localization system for mobile robots running in outdoor environments where GPS data is sometimes not accurate enough. Our system is equipped with GPS, encoder and magnetic sensor. Data obtained by these sensors are integrated using Kalman filter. GPS observation is omitted when the robot is running through particular areas. These areas are determined based on standard deviation of the data observed in advance.

Face detection algorithm used in Adaboost and Haar-like features were detected using the kana-kanji area detection. Create a set of images for learning. It was made by changing the number of sheets of the correct answer image and non-correct answer image. Correct answer images has Grayscale images and binary font images. Font used MS Gothic and MS Mincho. Non-correct answer images using gray-scale images of characters not included. The detection algorithm was verified by using 15 images for the verification. Verification images used 11 images including characters, two images including correct answer images, and two images not including characters. Adaboost and Haar-like feature using kana-kanji detection algorithms were validated.

Saffron pistils have been used, for example, as spice, dye and medicine from old days. But saffron pistils are considerably expensive, and are regarded to be luxury food. Saffron pistil is normally harvested by hand now. However, as the amount of pistils of one flower is very small, task of harvesting them is very hard. Therefore, in this paper, we discuss a method to automate harvesting saffron pistil using robot. We examine technique to detect cutting point and gripping point which are necessary for harvesting by online image processing. We verified the feasibility of our method through experiments.

For multi-object tracking using a particle filter, a tracking method in which Cross entropy is incorporated into a likelihood function is proposed, with the aim of improving the tracking speed. Baseline methods have utilized Bhattacharyya distance, KL divergence, and so on, in the likelihood function. However, these methods require unnegligible computational cost in calculation of color histograms for each sample, drawn at each frame. In contrast, in the Cross entropy method, likelihood calculations can be performed without generating sample histograms, which is expected to speed-up the tracking speed. Moreover, incorporating the background information into a tracking algorithm is a possible solution for performance improvement. Background information can be utilized together with cross entropy without increasing the computational cost. Therefore, fast and robust tracking algorithm for occlusion problem can be generated by incorporating background information with cross entropy. The proposed method was experimentally compared with a baseline method using the Bhattacharyya distance. The effectiveness of the proposed method and the effect of the number of sample were examined. © 2010, The Institute of Image Electronics Engineers of Japan. All rights reserved.

Recently, cleaning robots for domestic use are developed. However, present algorithm cannot always sweep the workspace uniformly, and this causes the inefficiency of the algorithm. The best sweeping algorithm may depend on the shape of the room. To select the best algorithm, we examined a method to acquire the shape of a room in the last report. In this report, we adopt an actual robot model so that the robot can distinguish rooms in the real world.

All stigmas of saffrons which are used as a spice or a dye are harvested by hand. However, it is heavy work to gather an enough amount. The purpose of this paper is to examine whether harvesting saffron's stigmas automatically is possible by using image processing. As a process of harvesting saffron stigmas automatically, we conceieved an idea of cutting the root of flower and harvesting stigma which is hanging by turning it down. Features to harvest stigmas are detected by using color information. The feasibility of this method was shown by the experiment.

The form of the putt swing is different among person. This suggests that an advice that can effectively improve a person's skill may differ between them. This paper describes a method to classify putt swing form aiming to develop an advisory system. We obtained motion data of subjects using a motion capture system. The putt swing form was appropriately classified using cluster analysis.

We propose a technique that automatically estimates a sports skill of examinees. In our study, we are discussing the technique for golf putting. However, we have estimated the skill only from apart of the examinee's motion data. It became the problem that the arbitrariness characteristics occur when choosing sensor. In this study, we introduce ensemble leaning to treat the whole motion data with Recurrent Neural Networks as temporary estimate sports skill and unify these result. The experimental result suggested the effectiveness of our method.

Recently, cleaning robots for domestic use have been actively developed. In our past studies, we introduced a new sweeping algorithm for cleaning robots, and verified its efficiency on a simulator. In this report, we mount the algorithm on a real robot, to compare the performance with the simulation result.

Nowadays, home-use robots that clean floors automatically are developed. However, the efficiency of these robots is not so good. That is because the best way of cleaning room depends on the room's shape. If the robot knows which room it is cleaning now, it will be possible to make a robot cleaning the room more efficiently. Therefore we propose a method which the robot goes around the room with hug motion, to detect the outline of the room.

We propose a visual-based online signature verification system in which the signer's pen tip is tracked. The input module of the system consists of only low-cost cameras (webcams) and does not need special equipment such as an electronic tablet. Online signature data is obtained from the images captured by the webcams by tracking the pen tip. The pen tip tracking is implemented by the Sequential Monte Carlo method in real time. Then, the distance between the input signature data and reference signature data enrolled in advance is computed. Finally, the input signature is classified as genuine or a forgery by comparing the distance with a threshold. lit this paper, we consider two camera positions. We performed an experiment using private data consisting of 250 signatures from 10 subjects. The experimental results show that the proposed system is promising for signature verification.

A novel algorithm is proposed for predicting transmembrane protein secondary structure from two-dimensional vector trajectories consisting of a hydropathy index and formal charge of a test amino acid sequence using stochastic dynamical system models. Two prediction problems are discussed. One is the prediction of transmembrane region counts another is that of transmembrane regions, i.e. predicting whether or not each amino acid belongs to a transmembrane region. The prediction accuracies, using a collection of well-characterized transmembrane protein sequences and benchmarking sequences, suggest that the proposed algorithm performs reasonably well. An experiment was performed with a glutamate transporter homologue from Pyrococcus horikoshii. The predicted transmembrane regions of the five human glutamate transporter sequences and observations based on the computed likelihood are reported. © Imperial College Press.

Many algorithms for online signature verification using multiple features have been proposed. Recently it has been argued that pen pressure, azimuth, and altitude can cause instability and deteriorate the performance. Algorithms without pen pressure and inclination features outperformed with them in SVC2004. However, we previously found that these features improved the performance in evaluations using our private database. The effectiveness of the features thus depended on the algorithm. Therefore, we re-evaluated our algorithm using the same database as used in SVC2004 and discuss the effectiveness of pen pressure, azimuth and altitude. Experimental results show that even though these features are riot so effective when they are used by themselves, they improved the performance when used in combination with other features. When pen pressure and inclination features were considered, an EER of 3.61% was achieved, compared to an EER of 5.79% when these features were not used.

A factor known as intersession variability in signatures causes deterioration of authentication performance. We propose a novel algorithm that includes a model updating scheme to correct for this variability. A model was provided for each user to calculate a score using fused multiple distance measures with respect to previous work. The algorithm consisted of an updating phase in addition to a training phase and a testing phase. In the training phase, the model's parameters were sampled using a Markov Chain Monte Carlo method for each individual. In the testing phase, the generated model was used to determine whether a test signature was genuine. In the updating phase, the parameters were updated with test data using a Sequential Monte Carlo (SMC) algorithm. Adoption of a hyper hyper parameter for automatically adjusting a hyper parameter in MC improved the authentication performance. Several experiments were performed on signatures from a public database. The proposed algorithm achieved an EER of 7.59%.

Personal authentication is becoming increasingly important. Biometrics, that is, the use of biological information, is one of the most promising techniques for this application. This paper proposes an online signature verification system. A serious problem in online signature verification is the difficulty of collecting enough signature data to generate a reliable model. In this paper, we propose a user-generic fusion model to resolve this problem. In the model generation, we use available datasets composed of genuine and forged signatures of many signers. The model's parameters are trained using the Markov Chain Monte Carlo method. We report experimental results of our proposed algorithm using two public databases.

Authentication of individuals is rapidly becoming an important issue. The authors previously proposed a Pen-input online signature verification algorithm. The algorithm considers a writer's signature as a trajectory of pen position, pen pressure, pen azimuth, and pen altitude that evolve over time, so that it is dynamic and biometric. Many algorithms have been proposed and reported to achieve accuracy for on-line signature verification, but setting the threshold value for these algorithms is a problem. In this paper, we introduce a user-generic model generated by AdaBoost, which resolves this problem. When user- specific models (one model for each user) are used for signature verification problems, we need to generate the models using only genuine signatures. Forged signatures are not available because imposters do not give forged signatures for training in advance. However, we can make use of another's forged signature in addition to the genuine signatures for learning by introducing a user generic model. And Adaboost is a well-known classification algorithm, making final decisions depending on the sign of the output value. Therefore, it is not necessary to set the threshold value. A preliminary experiment is performed on a database consisting of data from 50 individuals. This set consists of western-alphabet-based signatures provide by a European research group. In this experiment, our algorithm gives an FRR of 1.88% and an FAR of 1.60%. Since no fine-tuning was done, this preliminary result looks very promising.

Authentication of individuals is rapidly becoming an important issue. The authors have previously proposed a pen-input online signature verification algorithm. The algorithm considers writer's signature as a trajectory of pen-position, pen-pressure and pen-inclination which evolves over time, so that it is dynamic and biometric. In our previous work, genuine signatures were separated from forgery signatures in a linear manner. This paper proposes a new algorithm which performs nonlinear separation using Bayesian MCMC (Markov Chain Monte Carlo). A preliminary experiment is performed on a database consisting of 1825 genuine signatures and 4117 skilled forgery signatures from fourteen individuals. FRR 0.81% and FAR 0.87% are achieved. Since no fine tuning was done, this preliminary result looks very promising.

Authentication of individuals is rapidly becoming an important issue. The authors have previously proposed a pen-input online signature verification algorithm. The algorithm considers writer's signature as a trajectory of pen-position, pen-pressure and peninclination which evolves over time, so that it is dynamic and biometric. In our previous work, genuine signatures were separated from forgery signatures in a linear manner. This paper proposes a new algorithm which performs nonlinear separation using Bayesian MCMC (Markov Chain Monte Carlo). A preliminary experiment is performed on a database consisting of 1852 genuine signatures and 3170 skilled forgery signatures from fourteen individuals. FRR 0.81% and FAR 0.87% are achieved. Since no fine tuning was done, this preliminary result looks very promising.

Authentication of individuals is rapidly becoming an important issue. This paper proposes a new nonlinear algorithm for pen-input on-line signature verification incorporating pen-position, pen-pressure and pen-inclinations trajectories. A preliminary experiment is performed on a database consisting of 1849 genuine signatures and 3174 skilled(dagger) forgery signatures from fourteen individuals. Since no fine tuning was done, this preliminary result looks very promising.

Authentication of individuals is rapidly becoming an important issue. On-line signature verification is one of the methods that use biometric features. This paper proposes a new HMM algorithm is for on-line signature verification. After preprocessing, input signature is discretized in a polar coordinate system. This particular discretization leads to a simple procedure for assigning initial state and state transition probabilities. This paper utilizes only pen position trajectories, no other information is used which makes the algorithm simple and fast. A preliminary experiment shows that the proposed algorithm appears to be promising.

A new algorithm is proposed for inferring the number of transmembrane regions of transmembrane proteins from two dimensional vector trajectories consisting of hydropathy index and charge of amino acids by stochastic dynamical system models. The prediction accuracy of a preliminary experiment is 94%. Since no fine-tuning is done, this appears encouraging.