Using implicit neural representations, we proposed a fusion method of data captured by LiDAR and camera.
This method uses the neural density field from LiDAR data to represent the geometric information.
As well as NeRF, we achieve sensor calibration and data fusion simultaneously
by learning the color field that is consistent with input camera images.
Pesticide spraying is essential for growing crops, but food safety and cost are significant issues.
We have, therefore, developed a system to detect and remove worms using a quadruped robot capable of moving over uneven terrain.
We developed a mobile scanning system for fastly and accurately capturing the 3D range data.
Our system consists of a LiDAR and a color camera.
While the laser scanner works for capturing 3D structures,
the color camera captures an image sequence.
The sensor motion is estimated robustly from a sensor-fused 2D/3D feature-tracking method,
which helps to reconstruct the structures from the scan lines accurately.
We developed a flying sensor system to capture 3D data aerially.
The system, consisting of a omni-directional laser scanner and a panoramic camera,
can be mounted under a mobile platform to achieve the aerial scanning with high resolution and accuracy.
Since the laser scanner often requires several minutes
to complete an omni-directional scan, the raw data is distorted seriously due to the unknown and uncontrollable
movement during the scanning period. Our approach recovers the sensor motion by utilizing the spacial
and temporal features extracted both from the image sequences and point clouds.
We proposed an event camera tracking method for object manipulation by robots.
The method employs a unified camera projection model for various lenses, and an error function and optimization method that takes motion blur into account.
This study proposes a learning-based method to detect four different types of edges in 2D images.
The network architecture is based on the SWIN Transformer and is capable of extracting fine edges by using Dice loss.
This study provides a spherical harmonics (SH) based fast structural representation (SH-FS)
in visual SLAM using sparse point clouds, which extracts the structure information from sparse points into single vector.
SH-FS was applied in conventional feature-based loop closing process.
Furthermore, a structure-aware loop closing method in visual SLAM was proposed to improve the robustness of SLAM systems.
To take advantage of the drone's wide field of view,
we developed a drone pose estimation system from the ground vehicle.
In this system, the relative pose is obtained by direct measurement by LiDAR and
indirect measurement of the camera's vanishing directions.
We proposed a concept to attach the SLAM-device onto a robot and quickly realize the robot navigation in 3D space.
The proposed method calibrates SLAM-device and robot itself by using the relative poses obtained by several robot movements and clarified the efficient ones for the calibration according to the DoF of the robot.
Furthermore, the relative pose is dynamically refined so that the contact between the environment and the robot maintains the geometric consistency.
We propose a non-learning depth completion method for a sparse depth map captured using a LiDAR sensor guided by a pair of stereo images.
The proposed selective stereo matching (SSM) method searches the most appropriate depth value for each image pixel from its neighborly projected LiDAR points based on an energy minimization framework.
This depth selection approach can handle any type of mis-projection.
Moreover, SSM has an advantage in terms of long-range depth accuracy because it directly uses the LiDAR measurement rather than the depth acquired from the stereo.
We proposed methods for estimating a dense depth map from a sparse LIDAR point cloud and images.
Our unsupervised approach is a real-time dense depth completion from sparse depth maps guided by a single image.
Our method generates smooth depth maps while preserving discontinuity between different objects.
The key idea is a Binary Anisotropic Diffusion Tensor (B-ADT)
which can eliminate smoothness constraint at intended positions and directions
by applying variational regularization.
Another approach relies on a directionally biased propagation of known depth to missing areas based on semantic segmentation.
Additionally, we classify different object boundaries as either occluded or connected
to limit the extent of the data propagation.
At the regions with inevitably missing point cloud data,
we depend on estimated depth using motion stereo.
We propose a real-time dense 3D mapping method for fisheye cameras without explicit rectification and undistortion.
We extend the conventional variational stereo method by constraining the correspondence search
along the epipolar curve using a trajectory field induced by camera motion.
We also propose a fast way of generating the trajectory field without increasing the processing time
compared to conventional rectified methods.
We proposed a method to generate robot motions for vision-based teleoperation systems.
Task Model performs recognition and transmission of human motions
and can simultaneously solve the issues of the structural differences
between the human and the humanoid robot and time delays.
We also developed a method of articulation modeling using an RGB-D sensor.
The articulation parameters are estimated by a fusion of hand motion and point-cloud alignment.