Yoji Kuroda

Online road surface analysis using laser remission value in urban environments

This paper describes the novel road surface analysis using reflectivity of a laser scanner in structured outdoor environments. The proposed approach makes estimation of road surface conditions robust by using information of remission value as reflectivity of a laser that much less depends on brightness of color or ambient lighting than passive camera. Our method can be applied to various structured outdoor environments by online estimating distributions of the remission value from the road surface. This article shows that the method is successfully verified with accuracy of approximately 99% at both (i) the testing course of the 2009 Real World Robot Challenge which is known as “Tsukuba Challenge” and (ii) our university campus.

Visual odometry with effective feature sampling for untextured outdoor environment

In this paper, we propose stereo vision based visual odometry with an effective feature sampling technique for untextured outdoor environment. In order to extract feature points in untextured condition, we divide an image into some sections and affect suitable processes for each section. This approach can also prevent concentration of feature points, and the influence with a moving object can be reduced. Robust motion estimation is attained using the framework of 3- point algorithm and RANdom SAmple Consensus (RANSAC). Moreover, the accumulation error is reduced by keyframe adjustment. We present and evaluate experimental results for our system in outdoor environment. Proposed visual odometry system can localize the robots position within 4% error in untextured outdoor environment.

Mobile robot localization using multiple observations based on place recognition and GPS

In this paper, we propose a mobile robot localization system using multiple observations, which show the robots global position. One of observations is GPS observation, the other is utilized an appearance based place recognition. Using GPS observations has still some challenging problems such as multipath and signal lost under the environments there is tall buildings nearby. It affects a significant error on localization. On the other hand, appearance based place recognition methods are efficient to recognize the robots global position. It becomes possible to use a scene database with global position information. However, it could fail to function properly in natural environments like a lawn grass or trees in a park. We solve these demerits of each observations by using these multiple observations. Our system uses not only multiple observations but also dead reckoning with Gyrodometry model. As a result, the proposed localization system have achieved robust localization. To verify the validity of proposed method, our experiments using 1600m outdoor course in different seasons were conducted.

Online road segmentation for urban complex environments

In this paper, we propose a novel approach to stable near and long range perception for various complex outdoor environments. Our techniques cope robustly with near-range road estimation using a laser scanner and long-range terrain classification using a color camera. Near-range road surface conditions are estimated by using information of remission value as reflectivity of a laser. We apply graph cut algorithm to grid map in order to estimate road region robustly also in complex environments where fallen leaves exist sparsely. Moreover, we propose superpixel-based terrain classification method which can give a good performance compared with pixel-based classification. Experimental results have shown that demonstrate a marked increase in long-range classification and near-range road estimation accuracy over standard methods.

Near-to-Far Self-Supervised Road Estimation for Complicated Environments

Abstract In this paper, we propose a novel approach to stable near and long range perception for various complex outdoor environments. Our techniques cope robustly with near-range road estimation using a laser scanner and long-range terrain classification using a color camera. Near-range road surface conditions are estimated by using information of remission value as reflectivity of a laser. We apply graph cut algorithm to grid map in order to estimate road region robustly also in complex environments where fallen leaves exist sparsely. Moreover, we propose superpixel-based terrain classification method which can give a good performance compared with pixel-based classification. Our approach has real-time processing. Experimental results have shown that demonstrate a marked increase in long-range classification and near-range road estimation accuracy over standard methods.

Effective strategy for autonomous navigation without prior knowledge in FastSLAM

This paper describes the efficient strategy for planning for autonomous mobile robot navigation using the information which is the resulting probabilistic distribution of position and map acquired by solving the SLAM. In order to estimate good robots position and map, we used a highly efficient variant of the grid based version of the FastSLAM algorithm. D* Lite algorithm for global path planning, which has the effective replanning at the partial cost field changed, was employed. Because the acquired map in the SLAM is also grid based which indicates the probabilistic existence of the obstacles in each grid, and SLAMs uncertain grid map is utilized to compute the cost field for path planning. In this research, it was proven that the mobile robot could carry out autonomous navigation in the outdoor field without prior information. This paper presented that the mobile robot reached the predefined goal with estimating good position and map simultaneously.

Localization independent of location based on place recognition and GPS observations

In this research, we propose a mobile robot localization system using multiple observations, which show the robots global position. One of observations is GPS observations, another is utilized an appearance based place recognition. Using GPS observations faces still some challenging problems such as multipath and signal lost under the environments there is tall buildings nearby. These are critical issues for achieving a high accuracy and stable localization. On the other hand, appearance based place recognition methods are efficient to recognize the robots global position. It becomes possible to use a scene database with global position information. However appearance based place recognition methods could fail to function properly in natural environments like a lawn grass or trees in a park. We solve these disadvantages of each observations by using these multiple observations. Our system uses not only multiple observations but also dead reckoning with Gyrodometry model. Therefore, proposed method localize a robot position robustly indoors or not. To verify the validity of proposed method, our experiments are conducted about 1600m outdoor course in different seasons and course through an indoor.

Online motion model parameter estimation using Augmented Kalman Filter and discriminative training

In this paper, we propose an online motion model parameter estimation method. To achieve accurate localization, accurate estimation of motion model parameters is needed. However, the true values of motion model parameters change sequentially according to alteration of surrounding environments. Therefore the online estimation is absolutely imperative. As a typical method to estimate motion model parameters sequentially, Augmented Kalman Filter (AKF) is there. AKF achieve parameter estimation through Kalman filtering algorithm. However, AKF has serious problems to be implemented in real robot operation. These problems are the accuracy of observation and the limitation to motion control of robots. To solve these problems and achieve accurate motion model parameter estimation, proposed method introduces discriminative training. The introduction of discriminative training increases the convergence performance and stability of parameter estimation through AKF. The proposal method achieves accurate motion model parameter estimation in real robot operation. This paper describes the efficiency of our technique through simulations and an outdoor experiment.

Mobile robot localization using appearance based place recognition

In this paper, we propose a mobile robot localization system using appearance based place recognition. Our system uses appearance based place recognition and dead reckoning with Gyrodometry model. We use the theory of the observation likelihood estimation introduced from Fast Appearance Based Mapping (FAB-MAP) for appearance based place recognition. Images collected by a robot produce Bag-of-Words representation and Chow Liu trees. This representation is based on quantized SURF detectors/descriptors. A certain place can be recognized when the image collected matches a reference image coming from the same place in scene database. In addition, problems of appearance based place recognition techniques are solved by using sensor fusion. Experimental results are presented in an outdoor environment. The proposed localization system can achieve the average robots position error within 2% without GPS in an outdoor environment.

Correcting angle of visual odometry system by fusing monocular and stereo methods in untextured dynamic environment

In this research, we propose state-of-the-art 6-Degrees-Of-Freedom (6DOF) visual odometry (VO) system which fused stereo VO with effective feature sampling, VO using normal vector information of a ground plane and monocular VO. Firstly, stereo VO is used to do 6DOF motion estimation in untextured dynamic environments. Secondly, ground plane information is used to improve the accuracy of roll and pitch angles estimation. Thirdly, monocular VO is introduced to solve the problems in stereo processing. Finally, Unscented Kalman Filter is adopted for the fusion of the information acquired from three kinds of VO techniques to accomplish robust and accurate localization. We present and evaluate experimental results for our system over 730 meters runs in a challenging outdoor environment and compare it with ground truth. Proposed VO system can localize the robots position within 3.2% error in untextured outdoor environment.