Ryunosuke Hamada

Vehicle detection and localization on bird's eye view elevation images using convolutional neural network

For autonomous vehicles, the ability to detect and localize surrounding vehicles is critical. It is fundamental for further processing steps like collision avoidance or path planning. This paper introduces a convolutional neural network- based vehicle detection and localization method using point cloud data acquired by a LIDAR sensor. Acquired point clouds are transformed into birds eye view elevation images, where each pixel represents a grid cell of the horizontal x-y plane. We intentionally encode each pixel using three channels, namely the maximal, median and minimal height value of all points within the respective grid. A major advantage of this three channel representation is that it allows us to utilize common RGB image-based detection networks without modification. The birds eye view elevation images are processed by a two stage detector. Due to the nature of the birds eye view, each pixel of the image represent ground coordinates, meaning that the bounding box of detected vehicles correspond directly to the horizontal position of the vehicles. Therefore, in contrast to RGB-based detectors, we not just detect the vehicles, but simultaneously localize them in ground coordinates. To evaluate the accuracy of our method and the usefulness for further high-level applications like path planning, we evaluate the detection results based on the localization error in ground coordinates. Our proposed method achieves an average precision of 87.9% for an intersection over union (IoU) value of 0.5. In addition, 75% of the detected cars are localized with an absolute positioning error of below 0.2m.

Two-stage Hybrid A* path-planning in large petrochemical complexes

In this study, we aim to achieve path-planning for firefighter robots in large petrochemical complexes. In large environments, path-planning (e.g., Hybrid A*) requires a large computation memory and a long execution time. These constrains are not feasible for firefighter robots. In order to overcome these two challenges, we propose a two-stage hybrid A* path-planning. For the first stage we use a global path-planner that makes a path using a low-resolution grid map of 2 m. The global path-planner generates a path for an area of approx. 500 m ×1000 m in 10 seconds. In the second stage, we refine the path by using a local-planner that uses a local-map of 100 m ×100 m size around the robot with a high resolution grid of 1 m. The local planner receives its sub-goal from the global planner and recalculates a local path at a high speed of a few hundred milliseconds. Therefore, the local-planner can react to changes of the map due to obstacles in real-time. We evaluated our proposed method by comparing with conventional hybrid A* in simulated as well as real experimental data of petrochemical complexes. By employing the local-planner our method could drastically reduce the used memory and execution time for the re-planning. For a trajectory of 600 m, our method reduces the execution time by 99.2% for real data and by 94.34% for simulated data. The memory usage was likewise drastically reduced by 97.45% for real data and by 97.91% for simulated data.

Generation of Turning Motion for Tracked Vehicles Using Reaction Force of Stairs’ Handrail

Inspections by mobile robots are required in chemical and steel plants. The robots are required to ascend and descend stairs because equipment components are installed on different-level floors. This paper proposes turning motion for tracked vehicles on stairs. A characteristic of the proposed turning motion is that it is generated using the reaction force from the safety wall of the stairs’ handrail. The safety wall is commonly used in plants because it prevents objects from dropping down and damaging equipments. Proper turning motion is generated based on the motion model of the tracked vehicle. Experimental results show that the proposed turning motion can change the heading direction on the stairs. In addition, the proposed turning motion enables the vehicle to run with less slippage, as compared to other turning motions. The proposed method can reduce slippage by 88% while climbing up the stairs and by 44% while climbing down the stairs. The proposed method is more effective on the upward stairs than on the downward stairs. An autonomous turning motion control is implemented on the tracked vehicle, and it is evaluated on the upward stairs.

Robust stairway-detection and localization method for mobile robots using a graph-based model and competing initializations

One of the major challenges for mobile robots in human-shaped environments is navigating stairways. This study presents a method for accurately detecting, localizing, and estimating the characteristics of stairways using point cloud data. The main challenge is the wide variety of different structures and shapes of stairways. This challenge is often aggravated by an unfavorable position of the sensor, which leaves large parts of the stairway occluded. This can be further aggravated by sparse point data. We overcome these difficulties by introducing a three-dimensional graph-based stairway-detection method combined with competing initializations. The stairway graph characterizes the general structural design of stairways in a generic way that can be used to describe a large variety of different stairways. By using multiple ways to initialize the graph, we can robustly detect stairways even if parts of the stairway are occluded. Furthermore, by letting the initializations compete against each other, we find the best initialization that accurately describes the measured stairway. The detection algorithm utilizes a plane-based approach. We also investigate different planar segmentation algorithms and experimentally compare them in an application-orientated manner. Our system accurately detects and estimates the stairway parameters with an average error of only 2 . 5 mm for a variety of stairways including ascending, descending, and spiral stairways. Our method works robustly with different depth sensors for either small- or large-scale environments and for dense and sparse point cloud data. Despite this generality, our system’s accuracy is higher than most state-of-the-art stairway-detection methods.

Evaluation of LIDAR and GPS based SLAM on fire disaster in petrochemical complexes

Firefighter robot autonomy is important for fire disaster response robotics. SLAM is a key technology for the autonomy. We want to know if SLAM can be used in fire disasters. However, evaluating SLAM in an actual fire disaster is not possible because we cannot generate large fires in actual petrochemical complexes. In this study, we simulated a fire disaster, collected sensor data for different conditions in the fire disaster, and evaluated the accuracy of the SLAM. The fire effect for LIDAR was analyzed and the effect embedded in the LIDAR measurement simulator. Several sensor interval parameters used by a heat protection cover was also analyzed for protecting sensor from heat. The evaluation result show the best parameter is 1 s measurement and 9 s sensor cooling which the average accuracy of GPS and LIDAR based SLAM was in the range 0.25 — 0.36 m in the most difficult scenario in the petrochemical complex, has dimensions 1000 m × 600 m. Using the simulator enables us to evaluate the best interval parameter of GPS and LIDAR based SLAM at the fire disaster. The knowledge from the fire effect of the LIDAR could be used to improve LIDAR measurement in actual fire disasters.

Search Activities in Outdoor Environments by Cyber Rescue Dogs with Recognition Assistance

This paper reports results on our verification experiments for outdoor disaster response using rescue dogs. We have developed environment recognition system that especially focuses on constructing recognition functions quickly. We embedded this result into the system of cyber rescue dog. We designed a scenario that assumed disaster-stricken situation, and confirmed the effectiveness of the system.

Stable Haptic Feedback Generation During Mid Air Interactions Using Hidden Markov Model Based Motion Synthesis

Generation of stable and realistic haptic feedback in 3 dimensional midair interaction systems has garnered significant research interests recently. But the limitations in the sensing technologies such as unstable tracking, range limitations and occlusions occurred during interactions, along with the motion recognition faults significantly distort motion based haptic feedback. In this paper, a Hidden Markov Model based motion element synthesis for stable haptic feedback generation is proposed. The subjective evaluation experimental results using the proposed model on 3 subjects during a zooming task have shown improvements in user perception of the gestures.