In a 3D scanning system, using a camera and a line laser, it is critical to obtain the exact geometrical relationship between the camera and laser for precise 3D reconstruction. With existing depth cameras, it is difficult to scan a large object or multiple objects in a wide area because only a limited area can be scanned at a time. We developed a 3D scanning system with a rotating line laser and wide-angle camera for large-area reconstruction. To obtain 3D information of an object using a rotating line laser, we must be aware of the plane of the line laser with respect to the camera coordinates at every rotating angle. This is done by estimating the rotation axis during calibration and then by rotating the laser at a predefined angle. Therefore, accurate calibration is crucial for 3D reconstruction. In this study, we propose a calibration method to estimate the geometrical relationship between the rotation axis of the line laser and the camera. Using the proposed method, we could accurately estimate the center of a cone or cylinder shape generated while the line laser was rotating. A simulation study was conducted to evaluate the accuracy of the calibration. In the experiment, we compared the results of the 3D reconstruction using our system and a commercial depth camera. The results show that the precision of our system is approximately 65% higher for plane reconstruction, and the scanning quality is also much better than that of the depth camera.

Head pose estimation (HPE) is a critical task for numerous applications ranging from human-computer interaction, healthcare, and robotics, to surveillance. Most existing methods employ Euler angles as a representation, which often face challenges such as a gimbal lock, especially in full-range rotation scenarios or rotation matrices that require nine parameters. This study introduces WQuatNet, a novel deep learning-based model that leverages the quaternion representation, which uses only four parameters, to avoid this challenge. WQuatNet was designed based on a landmark-free HPE method to predict head poses across the full-range angles of 360\(^{\circ }\) from images. Landmark-free methods bypass the need for explicit detection of facial landmarks; instead, they leverage the entire image to estimate the head orientation. The model incorporates a RepVGG-D2se backbone for robust feature extraction and introduces two loss functions tailored for quaternion predictions. Our experimental results on multiple HPE datasets covering both narrow- and full-range angles demonstrate that WQuatNet outperforms the state-of-the-art (SOTA) approaches in terms of accuracy. The performance of the proposed HPE was evaluated using the CMU, AGORA, BIWI, AFLW2000, and 300W-LP datasets. We also perform ablation studies and error analyses to validate the significance of each component of the model.

Although non-invasive brain stimulation techniques do not involve surgical procedures, the challenge remains in correctly locating the stimulator from outside the head. There is a limit to which one can manually and precisely position and orient the stimulator or repeatedly move the stimulator around the same position. Therefore, in this study, we developed a serial robot with 6 degrees-of-freedom to move the stimulator and a neuro-navigation system to determine the stimulus point from looking at the shape of the subject’s brain. The proposed robot applied a spherical mechanism while considering the safety of the subject, and the workspace of the robot was designed considering the shape of the human head. Position-based visual servoing was applied to compensate for unexpected movements during subject stimulation. We also developed a neuro-navigation system that allows us visually to check the focus of the stimulator and the human brain at the same time and command the robot to the desired point. To verify the system performance, we first performed repeatability and motion compensation experiments of the robot and then evaluated the repeated biosignal response experiments through transcranial magnetic stimulation, a representative technique of non-invasive brain stimulation.

Ultrasonic neuromodulation (UNMOD) provides a non-invasive brain stimulation. However, the high-resolution region-specificity of UNMOD with a single element transducer combined with a mechanical positioning system could have limits due to the intrinsic positioning error from mechanical systems. A phased array system could lead to highly selective neuromodulation with electronic control. A specialized phased-array system with a robotic arm is implemented for a rhesus monkey model. Various primary motor cortex areas related to tail, hand, and mouth were stimulated with a 200 μm step size. The ultrasonic parameters were I_SPTA of 840 mW/cm², pulse repetition frequency of 100 Hz, and a 5% duty factor at 600 kHz. The induced movement were recorded and analyzed. Separate digits, mouth, and tongue motions were successfully induced by electronically controlling the focus. The identical body part movement could be induced when the focus was moved back to the identical primary motor cortex with electronic control. Accordingly, the reproducibility of UNMOD could be partially validated with rhesus monkey model. A phased-array system appears to have a potential for the non-invasive and region-selective neuromodulation method.

Real-time 6D pose tracking using camera images plays an important role in enabling robots to track and grasp household items. Although iterative closest point (ICP) is frequently employed as the primary method for pose tracking owing to its good performance and low resource consumption, there are several challenges such as the difficulty in dealing with flat or symmetric objects. These challenges make the ICP prone to failure in tracking scenarios. In this study, we introduce an ICP enhancement algorithm to effectively utilize the ICP for tracking household items. Our main contribution can be summarized in three ways: usage of RGB keypoints for obtaining initial pose, refinement approach for dealing with flat or rotationally symmetric objects, and self-recovery algorithm under tracking failure. We estimate the initial pose using RGB keypoints and fine-tune the weights of the loss function to achieve an accurate pose by considering the flat and rotational symmetric properties of the object. Additionally, a recovery algorithm with the Oriented FAST and Rotated BRIEF (ORB) descriptor was employed to detect and resolve tracking failures. While the performance was relatively limited for objects with low RGB texture, our algorithm achieved stable and high accuracy when sufficient RGB texture was present. We confirmed that our algorithm achieved 100% ADDs-AUC on 16 objects in YCB Video dataset.