Takuma Ito
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Takuma Ito has a total of 45 co-authors. Here are the top co-authors by number of shared publications:
Minoru Kamata
The University of Tokyo, Tokyo, Japan
18 shared publicationsKyoichi Tohriyama
Toyota Motor Corporation, Shizuoka, Japan
6 shared publicationsTatsuya Shino
The University of Tokyo
5 shared publicationsAkito Yamasaki
Meijo University, Nagoya, Japan
4 shared publicationsHideo Inoue
Kanagawa Institute of Technology, Kanagawa, Japan
4 shared publicationsKota Watanabe
4 shared publications
Latest Publications
- Open Access
Position Estimation Considering Uncertain Classification of Cyclists Based on Partially Observed Movement Characteristics
Prevention of crossing collisions between cyclists and vehicles at nonsignalized intersections on community roads where walls and hedges limit visibility is required in Japan. Because available observation information in real-time is limited on community roads, the use of statistical information that represents the typical movement characteristics of cyclists is effective to compensate for the lack of observation information. From such a background, in our previous study, we proposed a method to construct “location-dependent statistical information” (LDSI) and a method to utilize it as “virtual observation” (VO) and “virtual control input” (VCI) in stochastic position estimation. Here, although LDSI was constructed for multiple clusters of cyclists, the classification method of the cyclists observed in real-time was not considered. In the real world, the limitation of the observation information causes classification uncertainty. Thus, in this study, we propose a position estimation method that utilizes soft classification results and considers classification uncertainty by integrating VO and VCI derived from LDSI of each cluster. To evaluate the proposed method in this study, we conduct a simulation and an experiment in the real world. Through the comparison with conventional methods, we confirm that our proposed method in this study improves the performance of the position estimation. The proposed method will contribute to a cooperative safety system.
Open AccessPosition Estimation Considering Uncertain Classification of Cyclists Based on Partially Observed Movement Characteristics
Prevention of crossing collisions between cyclists and vehicles at nonsignalized intersections on community roads where walls and hedges limit visibility is required in Japan. Because available observation information in real-time is limited on community roads, the use of statistical information that represents the typical movement characteristics of cyclists is effective to compensate for the lack of observation information. From such a background, in our previous study, we proposed a method to construct “location-dependent statistical information” (LDSI) and a method to utilize it as “virtual observation” (VO) and “virtual control input” (VCI) in stochastic position estimation. Here, although LDSI was constructed for multiple clusters of cyclists, the classification method of the cyclists observed in real-time was not considered. In the real world, the limitation of the observation information causes classification uncertainty. Thus, in this study, we propose a position estimation method that utilizes soft classification results and considers classification uncertainty by integrating VO and VCI derived from LDSI of each cluster. To evaluate the proposed method in this study, we conduct a simulation and an experiment in the real world. Through the comparison with conventional methods, we confirm that our proposed method in this study improves the performance of the position estimation. The proposed method will contribute to a cooperative safety system.
- Open Access
- 16 April 2026
- Published by Springer Nature in International Journal of Intelligent Transportation Systems Research
- Vol. 24 (2) , 1172-1189
Visibility Evaluation Around Unsignalized Intersections with Occlusion From Arbitrary On-road Viewpoints Using Point Cloud Maps
To prevent traffic accidents at unsignalized intersections, it is essential to proactively evaluate potential risks around intersections. Although existing research has developed a method to evaluate the visibility around intersections from the vehicle driver’s viewpoint for risk evaluation, the visibility from that of Vulnerable Road Users (VRUs) is not assumed. Thus, this study proposes a framework for evaluating visibility from arbitrary viewpoints on the road, considering all types of traffic participants. As an indicator related to visibility evaluation, the framework uses the Field-of-View (FOV), which represents visibility around the intersection. To generate an FOV map, the framework constructs obstacle points from a point cloud map generated by a LiDAR SLAM algorithm. Then, comprehensive viewpoints are placed along the road network, and the FOV is calculated from each viewpoint. The proposed framework is implemented and tested in real-world environments. The evaluation results show that the mean absolute error between the calculated and measured FOV is 2.06 deg, which is under the estimation resolution required to distinguish a pedestrian on the intersecting road, indicating sufficient precision for practical use in future safety applications. Furthermore, the framework correctly evaluated variations in the FOV caused by differences in geometrical characteristics of intersections, travel positions on the road, and viewpoint heights.
- 16 April 2026
- Published by Springer Nature in International Journal of Intelligent Transportation Systems Research
- Vol. 24 (2) , 1172-1189
Open AccessVisibility Evaluation Around Unsignalized Intersections with Occlusion From Arbitrary On-road Viewpoints Using Point Cloud Maps
To prevent traffic accidents at unsignalized intersections, it is essential to proactively evaluate potential risks around intersections. Although existing research has developed a method to evaluate the visibility around intersections from the vehicle driver’s viewpoint for risk evaluation, the visibility from that of Vulnerable Road Users (VRUs) is not assumed. Thus, this study proposes a framework for evaluating visibility from arbitrary viewpoints on the road, considering all types of traffic participants. As an indicator related to visibility evaluation, the framework uses the Field-of-View (FOV), which represents visibility around the intersection. To generate an FOV map, the framework constructs obstacle points from a point cloud map generated by a LiDAR SLAM algorithm. Then, comprehensive viewpoints are placed along the road network, and the FOV is calculated from each viewpoint. The proposed framework is implemented and tested in real-world environments. The evaluation results show that the mean absolute error between the calculated and measured FOV is 2.06 deg, which is under the estimation resolution required to distinguish a pedestrian on the intersecting road, indicating sufficient precision for practical use in future safety applications. Furthermore, the framework correctly evaluated variations in the FOV caused by differences in geometrical characteristics of intersections, travel positions on the road, and viewpoint heights.
- Open Access
V2N-Based Comprehensive Safety Framework by Prediction of VRU Movement on Community Roads with Management of Route Branching at Intersections
Traffic accidents involving Vulnerable Road Users (VRUs) frequently occur at unsignalized intersections on Japanese community roads. To prevent such accidents, collision avoidance systems need to predict VRUs’ movements throughout the entire road network while explicitly handling uncertainty degraded by sparse observations and frequent route branching at intersections. Based on this motivation, this study proposes a Vehicle-to-Network (V2N)-based comprehensive safety framework for estimation of VRU movement and prediction of future intersection entry for community roads. The framework integrates estimation results provided from Roadside Edges and Vehicle Edges at a Central Server. In addition, road geometry from map information is incorporated as pseudo-observations into the estimation, and multiple route hypotheses are explicitly managed to represent route branching at intersections. For intersection-entry prediction, entry certainty is calculated by integrating a predicted distribution. For evaluation of the proposed framework, we conduct Monte Carlo simulations on simplified grid road networks. The results show that the proposed framework maintains conservative estimation under sparse observations and improves prediction when additional observation information from surrounding vehicles becomes available. Furthermore, a simulation-based case study using an actual community road-network geometry shows the feasibility of the proposed framework for cooperative collision avoidance on actual community roads.
Open AccessV2N-Based Comprehensive Safety Framework by Prediction of VRU Movement on Community Roads with Management of Route Branching at Intersections
Traffic accidents involving Vulnerable Road Users (VRUs) frequently occur at unsignalized intersections on Japanese community roads. To prevent such accidents, collision avoidance systems need to predict VRUs’ movements throughout the entire road network while explicitly handling uncertainty degraded by sparse observations and frequent route branching at intersections. Based on this motivation, this study proposes a Vehicle-to-Network (V2N)-based comprehensive safety framework for estimation of VRU movement and prediction of future intersection entry for community roads. The framework integrates estimation results provided from Roadside Edges and Vehicle Edges at a Central Server. In addition, road geometry from map information is incorporated as pseudo-observations into the estimation, and multiple route hypotheses are explicitly managed to represent route branching at intersections. For intersection-entry prediction, entry certainty is calculated by integrating a predicted distribution. For evaluation of the proposed framework, we conduct Monte Carlo simulations on simplified grid road networks. The results show that the proposed framework maintains conservative estimation under sparse observations and improves prediction when additional observation information from surrounding vehicles becomes available. Furthermore, a simulation-based case study using an actual community road-network geometry shows the feasibility of the proposed framework for cooperative collision avoidance on actual community roads.
- Open Access
Correction: Suzuki, K.; Ito, T. Virtual Observation Using Location-Dependent Statistical Information of Cyclists’ Movement for Estimation of Position and Uncertainty. Sensors 2025, 25, 5122
An implementation mistake was found in our simulation program which makes minor changes in some figures and a table in the original publication
Open AccessCorrection: Suzuki, K.; Ito, T. Virtual Observation Using Location-Dependent Statistical Information of Cyclists’ Movement for Estimation of Position and Uncertainty. Sensors 2025, 25, 5122
An implementation mistake was found in our simulation program which makes minor changes in some figures and a table in the original publication
- Open Access
- 1 January 2026
- Published by Society of Automotive Engineers of Japan, Inc. in International Journal of Automotive Engineering
- Vol. 17 (1) , 28-42
Method of Improving Machine Learning Model for Speed Decision on Community Roads by Extracting Potentially Dangerous Spots
Since traffic accidents on Japanese community roads are one of unsolved problems, we focus on a speed decision machine learning model (SDMLM) targeting community roads. On community roads, occlusions caused by housing walls make invisible risks of collisions at some spots, which we call “potentially dangerous spots.” Although SDMLM needs to output a safe speed at the potentially dangerous spots, SDMLM that does not learn from driving data including occlusionrelated risk might be unable to calculate a suitable speed. Thus, we develop a method for improving the SDMLM by utilizing careful driving data. In the improvement, the trade-off between “adaptation to occlusion-related risk” and “prevention of overcautiousness” needs to be considered. Therefore, for training an improved SDMLM, we extract potentially dangerous spots and prepare careful driving data at extracted spots. Through evaluations using the driving data in the real world, we confirm the effectiveness of the proposed method for improving the SDMLM on community roads. The results of this study will contribute to the development of the component of autonomous driving systems for preventing traffic accidents on Japanese community roads.
- 1 January 2026
- Published by Society of Automotive Engineers of Japan, Inc. in International Journal of Automotive Engineering
- Vol. 17 (1) , 28-42
Open AccessMethod of Improving Machine Learning Model for Speed Decision on Community Roads by Extracting Potentially Dangerous Spots
Since traffic accidents on Japanese community roads are one of unsolved problems, we focus on a speed decision machine learning model (SDMLM) targeting community roads. On community roads, occlusions caused by housing walls make invisible risks of collisions at some spots, which we call “potentially dangerous spots.” Although SDMLM needs to output a safe speed at the potentially dangerous spots, SDMLM that does not learn from driving data including occlusionrelated risk might be unable to calculate a suitable speed. Thus, we develop a method for improving the SDMLM by utilizing careful driving data. In the improvement, the trade-off between “adaptation to occlusion-related risk” and “prevention of overcautiousness” needs to be considered. Therefore, for training an improved SDMLM, we extract potentially dangerous spots and prepare careful driving data at extracted spots. Through evaluations using the driving data in the real world, we confirm the effectiveness of the proposed method for improving the SDMLM on community roads. The results of this study will contribute to the development of the component of autonomous driving systems for preventing traffic accidents on Japanese community roads.
- Open Access
- 10 December 2025
- Published by Springer Nature in International Journal of Intelligent Transportation Systems Research
- Vol. 24 (2) , 280-300
Placement Design of Various Roadside Sensors for Stochastic Position Prediction of Traffic Participants on Community Roads
In Japan, traffic accidents on community roads where visibility is limited at non-signalized intersections are one of the social issues. Although cooperative systems using roadside sensors are one of the promising solutions to this issue, cost constraints require strategic approaches to the placement of limited number of sensors. For the above reasons, this paper proposes a framework for exploring cost-efficient placement of Roadside Units (RSUs). To be more precise, this study focuses on the investigation of placement and combination of three types of sensors, which include cameras, radars, and LiDARs, in a 3 × 3 square grid town with different length of road segments. Then, we utilize Monte Carlo simulations to evaluate performance of position prediction of traffic participants approaching to intersections. After that, we explore placements by Non-dominated Sorting Genetic Algorithm II (NSGA-II). Through the proposed framework, we obtain diverse Pareto-optimal solutions which show a trade-off between the lowness of the total sensor costs and the performance of the position prediction. In addition, we obtained cost-efficient placement strategies for different road networks by analysis of Pareto-optimal solutions. This study offers a cost-efficient and adaptable framework for the placement of RSUs, promising enhanced cooperative systems and improved road safety.
- 10 December 2025
- Published by Springer Nature in International Journal of Intelligent Transportation Systems Research
- Vol. 24 (2) , 280-300
Open AccessPlacement Design of Various Roadside Sensors for Stochastic Position Prediction of Traffic Participants on Community Roads
In Japan, traffic accidents on community roads where visibility is limited at non-signalized intersections are one of the social issues. Although cooperative systems using roadside sensors are one of the promising solutions to this issue, cost constraints require strategic approaches to the placement of limited number of sensors. For the above reasons, this paper proposes a framework for exploring cost-efficient placement of Roadside Units (RSUs). To be more precise, this study focuses on the investigation of placement and combination of three types of sensors, which include cameras, radars, and LiDARs, in a 3 × 3 square grid town with different length of road segments. Then, we utilize Monte Carlo simulations to evaluate performance of position prediction of traffic participants approaching to intersections. After that, we explore placements by Non-dominated Sorting Genetic Algorithm II (NSGA-II). Through the proposed framework, we obtain diverse Pareto-optimal solutions which show a trade-off between the lowness of the total sensor costs and the performance of the position prediction. In addition, we obtained cost-efficient placement strategies for different road networks by analysis of Pareto-optimal solutions. This study offers a cost-efficient and adaptable framework for the placement of RSUs, promising enhanced cooperative systems and improved road safety.
- Open Access
Robust Pose Estimation and Size Classification for Unknown Dump Truck Using Normal Distribution Transform
Labor shortage has been a severe problem in the Japanese construction industry, and the automation of construction work has been in high demand. One of the needs is the automation of soil loading onto dump trucks. This task requires pose estimation and size classification of the dump trucks to determine the appropriate loading position and volume. At actual construction sites, specifications of dump trucks are not always known in advance. However, most of the existing methods cannot robustly estimate the pose and the size of such unknown dump trucks. To address this issue, we propose a two-stage method that estimates the pose of dump trucks and then classifies their size categories. We use Normal Distribution Transform (NDT) for pose estimation of dump trucks. Specifically, we utilize NDT templates of dump trucks which distinguish global differences among size categories and simultaneously absorb local shape variations within each category. The proposed method is evaluated by data in a real-world environment. The proposed method appropriately estimates the pose of dump trucks under various settings of positions and orientations. In addition, the method correctly classifies the observed dump truck with all three predefined size categories. Furthermore, the computation time is approximately 0.13 s, which is sufficiently short for practical operation. These results indicate that the method will contribute to the automation of soil loading onto dump trucks with unknown specifications.
Open AccessRobust Pose Estimation and Size Classification for Unknown Dump Truck Using Normal Distribution Transform
Labor shortage has been a severe problem in the Japanese construction industry, and the automation of construction work has been in high demand. One of the needs is the automation of soil loading onto dump trucks. This task requires pose estimation and size classification of the dump trucks to determine the appropriate loading position and volume. At actual construction sites, specifications of dump trucks are not always known in advance. However, most of the existing methods cannot robustly estimate the pose and the size of such unknown dump trucks. To address this issue, we propose a two-stage method that estimates the pose of dump trucks and then classifies their size categories. We use Normal Distribution Transform (NDT) for pose estimation of dump trucks. Specifically, we utilize NDT templates of dump trucks which distinguish global differences among size categories and simultaneously absorb local shape variations within each category. The proposed method is evaluated by data in a real-world environment. The proposed method appropriately estimates the pose of dump trucks under various settings of positions and orientations. In addition, the method correctly classifies the observed dump truck with all three predefined size categories. Furthermore, the computation time is approximately 0.13 s, which is sufficiently short for practical operation. These results indicate that the method will contribute to the automation of soil loading onto dump trucks with unknown specifications.
- Open Access
Virtual Observation Using Location-Dependent Statistical Information of Cyclists’ Movement for Estimation of Position and Uncertainty
Crossing collisions between cyclists and automobiles around nonsignalized intersections on community roads, where visibility around the intersection is poor due to occlusions caused by house walls, is a social issue related to traffic safety in Japan. Because available observation information for collision prevention is limited on community roads, utilizing the accumulated data is useful to compensate for the lack of observation information. Given these motivations, we propose a movement estimation method of cyclists by combining information from roadside sensors with location-dependent statistical information. First, we develop a method for analyzing the location-dependent statistical information of cyclists on a certain road from accumulated GNSS data using the Kalman smoother. Then, we develop a method for stochastically predicting the movement of cyclists even outside the observation range of a roadside sensor by using the concept of “virtual observation” based on location-dependent statistical information. To evaluate the proposed method, we conduct an experiment to accumulate GNSS data from cyclists using smartphones. As a result of comparison with a conventional method, we confirm that our proposed method can reduce the uncertainty of the estimated position; further, the reduction in the uncertainty will contribute to traffic safety by future advanced driver assistance systems.
Open AccessVirtual Observation Using Location-Dependent Statistical Information of Cyclists’ Movement for Estimation of Position and Uncertainty
Crossing collisions between cyclists and automobiles around nonsignalized intersections on community roads, where visibility around the intersection is poor due to occlusions caused by house walls, is a social issue related to traffic safety in Japan. Because available observation information for collision prevention is limited on community roads, utilizing the accumulated data is useful to compensate for the lack of observation information. Given these motivations, we propose a movement estimation method of cyclists by combining information from roadside sensors with location-dependent statistical information. First, we develop a method for analyzing the location-dependent statistical information of cyclists on a certain road from accumulated GNSS data using the Kalman smoother. Then, we develop a method for stochastically predicting the movement of cyclists even outside the observation range of a roadside sensor by using the concept of “virtual observation” based on location-dependent statistical information. To evaluate the proposed method, we conduct an experiment to accumulate GNSS data from cyclists using smartphones. As a result of comparison with a conventional method, we confirm that our proposed method can reduce the uncertainty of the estimated position; further, the reduction in the uncertainty will contribute to traffic safety by future advanced driver assistance systems.
- Open Access
- 1 January 2025
- Published by Society of Automotive Engineers of Japan, Inc. in International Journal of Automotive Engineering
- Vol. 16 (4) , 96-111
Sensor Coverage Aware Probabilistic Data Association to Track Multiple Traffic Participants Using Sparsely Placed Roadside Sensors
Traffic accidents on community roads, which frequently have intersections with poor visibility, are one of the social issues in Japan. Although safety technologies that utilize roadside sensors are expected to be effective for Japanese community roads, only a few roadside sensors and limited sensor coverage are available on community roads in practice. In such an environment, it is difficult to consistently track multiple traffic participants by associating sensor observations of them from one sensor coverage to another coverage. To address this difficulty, we propose a data association method for multi-target tracking on the assumption that targets can be outside the sensor coverage. The proposed method calculates the existence probability of each target being within the sensor coverage at each time step and incorporates it as a prior probability in the data association process. In the simulation experiments, comparisons with existing methods demonstrate that the proposed method achieves a higher association success rate in various conditions. Furthermore, real-world experiments validate the feasibility of the proposed method.
- 1 January 2025
- Published by Society of Automotive Engineers of Japan, Inc. in International Journal of Automotive Engineering
- Vol. 16 (4) , 96-111
Open AccessSensor Coverage Aware Probabilistic Data Association to Track Multiple Traffic Participants Using Sparsely Placed Roadside Sensors
Traffic accidents on community roads, which frequently have intersections with poor visibility, are one of the social issues in Japan. Although safety technologies that utilize roadside sensors are expected to be effective for Japanese community roads, only a few roadside sensors and limited sensor coverage are available on community roads in practice. In such an environment, it is difficult to consistently track multiple traffic participants by associating sensor observations of them from one sensor coverage to another coverage. To address this difficulty, we propose a data association method for multi-target tracking on the assumption that targets can be outside the sensor coverage. The proposed method calculates the existence probability of each target being within the sensor coverage at each time step and incorporates it as a prior probability in the data association process. In the simulation experiments, comparisons with existing methods demonstrate that the proposed method achieves a higher association success rate in various conditions. Furthermore, real-world experiments validate the feasibility of the proposed method.
- Open Access
- 1 January 2025
- Published by Society of Automotive Engineers of Japan, Inc. in International Journal of Automotive Engineering
- Vol. 16 (3) , 54-65
Virtual Observation Using Statistical Information of Cyclist's Velocity for Estimation of Position and Uncertainty
Cooperative safety systems which utilize roadside sensors or smartphones are expected to be solutions for cyclistrelated accidents at intersections with poor visibility. However, in actual environments, it is required to estimate positions of traffic participants with the limited number of real-time sensors. To compensate for the lack of real-time sensor data, this study proposes a method to utilize statistical information as virtual observations for Kalman Filter. First, we design virtual observations based on the statistical data of cyclist's velocity and the time-series effect of position estimation. Then, through the evaluations in the simulations and the real-world experiments, we confirm that the uncertainty of position estimation is reduced by the proposed method. This study will contribute to the cooperative safety systems under the situation of the limited real-time sensor data.
- 1 January 2025
- Published by Society of Automotive Engineers of Japan, Inc. in International Journal of Automotive Engineering
- Vol. 16 (3) , 54-65
Open AccessVirtual Observation Using Statistical Information of Cyclist's Velocity for Estimation of Position and Uncertainty
Cooperative safety systems which utilize roadside sensors or smartphones are expected to be solutions for cyclistrelated accidents at intersections with poor visibility. However, in actual environments, it is required to estimate positions of traffic participants with the limited number of real-time sensors. To compensate for the lack of real-time sensor data, this study proposes a method to utilize statistical information as virtual observations for Kalman Filter. First, we design virtual observations based on the statistical data of cyclist's velocity and the time-series effect of position estimation. Then, through the evaluations in the simulations and the real-world experiments, we confirm that the uncertainty of position estimation is reduced by the proposed method. This study will contribute to the cooperative safety systems under the situation of the limited real-time sensor data.