Joongjin Kook

Detecting far-range cyclists from LiDAR point clouds is challenging because cyclists are small road users and are often represented by only a limited number of points at long distances. This study proposes a training-time object-level temporal merge teacher that uses frame-level object IDs in the Waymo Open Dataset to improve single-frame 3D detection of sparse far-range cyclists while preserving the inference-time efficiency of PointPillars. Current-frame cyclist objects are defined as sparse/far targets when their ground-truth boxes are at least 40 m from the ego vehicle and contain no more than 20 LiDAR points. During teacher construction, previous-frame cyclist points with the same object ID are aligned to the current object coordinate system and used to construct the teacher input. The student detector receives only the original single-frame point cloud during inference; therefore, previous frames and ground-truth object IDs are not required at deployment. In validation experiments, the F2 temporal-merge teacher improved cyclist L2 APH from 0.1500 to 0.1903 over the baseline, and the F2 KD student achieved cyclist L2 AP/APH scores of 0.3119/0.1906. The results indicate that the temporal-merge teacher, pair-based fine-tuning, and KD signals can jointly improve cyclist AP for a single-frame detector. However, because the no-KD control also outperformed the baseline, the observed improvement should be interpreted as the combined effect of the temporal teacher construction and the training pipeline rather than the isolated effect of the KD loss alone.

This paper presents a non-learning-based, seed-dependent, semi-automatic pedestrian candidate generation pipeline for LiDAR point clouds. The proposed method is designed to support 3D annotation workflows by reducing irrelevant candidate clusters while improving the reliability of pedestrian candidate selection under distance-dependent point sparsity. The pipeline integrates distance-aware DBSCAN clustering, Single Template Matching (STM), and Centralized Point Augmentation (CPA). First, LiDAR points within the camera field of view are preprocessed, and pedestrian candidate clusters are generated using DBSCAN parameters configured according to distance intervals. Ground-snapping-based bounding-box refinement and height-based filtering are then applied to improve geometric consistency and reduce non-pedestrian candidates. In the second stage, STM compares PCA-aligned projected silhouettes of candidate clusters with a seed pedestrian template to suppress false positives. To address silhouette instability caused by sparse mid-range pedestrian points, CPA adds centroid-contracted points in the projection-relevant plane before template matching. Experiments on pedestrian-containing frames from the KITTI dataset show that STM improves precision from 27.6% to 60.5% and increases the F1-score from 36.8% to 51.4% compared with the initial DBSCAN-based candidate generation stage. The final CPA configuration improves recall from 44.7% to 46.7% and the overall F1-score from 51.4% to 52.1%, while revealing a precision–recall trade-off. Supplementary IoU analysis shows that the final DBSCAN–STM–CPA configuration maintains meaningful spatial overlap with pedestrian ground-truth boxes, achieving 88.9% at 3D IoU ≥ 0.10 and 81.6% at BEV IoU ≥ 0.25. Runtime analysis further shows that height-based filtering reduces the average per-frame processing time from 151.5 ms to 125.1 ms, while the final CPA configuration introduces only a small overhead, resulting in 126.2 ms per frame. These results demonstrate that the proposed DBSCAN–STM–CPA pipeline can provide reliable pedestrian candidates for semi-automatic 3D labeling without requiring class-specific detector training.

This study proposes a novel heart rate estimation method that detects subtle cardiac-induced vibrations propagated through the cardiovascular system based on the ballistocardiography (BCG) principle, using a six-axis heart rate sensing device that integrates a three-axis accelerometer and a three-axis gyroscope. To validate the effectiveness of the proposed method, a comparative analysis was conducted against heart rate measurements obtained from photoplethysmography (PPG) sensors, which are widely used in conventional heart rate monitoring. Experiments were conducted on 20 adult participants, and frequency domain analysis was performed using different time windows of 30 s, 20 s, 8 s, and 4 s. The results showed that the 4 s window provided the highest accuracy in heart rate estimation, demonstrating that the proposed method can effectively capture fine cardiac-induced vibrations. This approach offers a significant advantage by utilizing inertial sensors commonly embedded in wearable devices for heart rate monitoring without the need for additional optical sensors. Compared to optical-based systems, the proposed method is more power-efficient and less affected by environmental factors such as ambient lighting conditions. The findings suggest that heart rate estimation using the six-axis heart rate sensing device presents a reliable, continuous, and non-invasive alternative for cardiovascular monitoring.

KSII Transactions on Internet and Information Systems TIIS Paper Details

In this paper, we propose a lightweight deep learning network (DRRU-Net) for image-splicing forgery detection. DRRU-Net is an architecture that combines RRU-Net for learning the visual content of images and image acquisition artifacts, and a JPEG artifact learning module for learning compression artifacts in the discrete cosine transform (DCT) domain. The backbone model of a network based on pre-training, such as CAT-Net, a representative network for image forgery detection, has a relatively large number of parameters, resulting in overfitting in a small dataset, which hinders generalization performance. Therefore, in this paper, the learning module is designed to learn the characteristics according to the DCT domain in real time without pre-training. In the experiments, the proposed network architecture and training method of DRRU-Net show that the network parameters are smaller than CAT-Net, the forgery detection performance is better than that of RRU-Net, and the generalization performance for various datasets can be improved.

Recently, musculoskeletal disorders (MSDs) caused by repetitive working postures in industrial sites have emerged as one of the biggest problems in the field of industrial health. The risk of MSDs caused by the repetitive working postures of workers is quantitatively evaluated by using NLE (NIOSH Lifting Equation), OWAS (Ovako Working-posture Analysis System), RULA (Rapid Upper Limb Assessment), REBA (Rapid Entire Body Assessment), etc. Methods used for the working posture analysis include vision-based analysis and motion capture analysis. Vision-based analysis is a method where an expert with ergonomics knowledge watches and manually analyzes recorded working images. Although the analysis is inexpensive, it takes a lot of time to analyze. In addition, the analyst’s subjective opinions or mistakes may be reflected in the results, so it may be somewhat unreliable. On the other hand, motion capture analysis can obtain more accurate and consistent results, but its measurement equipment is very expensive and it requires a large space for measurement. In this paper, we propose a computer-based automated REBA system that can evaluate, automatically and consistently, working postures in order to supplement the shortcomings of these existing methods. The CREBA system uses the body detection learning model of MediaPipe to detect the worker’s area in the recorded images and sets the body area based on the position of the face, detected using the face tracking learning model. In the set area, the positions of joints are tracked using the posture tracking learning model, and the angles of joints are calculated based on the joint positions using the inverse kinematics, and then by automatically calculating the degree of load of the working posture with the REBA evaluation method. In order to verify the accuracy of the evaluation results of the CREBA system, we compared them with the experts’ vision-based REBA evaluation results. The result of the experiment showed a slight difference of about 1.0 points between the evaluation results of the expert group and those of the CREBA system. It is expected that the ergonomic analysis method for the working posture used in this study will reduce workers’ labor intensity and improve their safety and efficiency.

SLAM technology, which is used for spatial recognition in autonomous driving and robotics, has recently emerged as an important technology to provide high-quality AR contents on mobile devices due to the spread of XR and metaverse technologies. In this paper, we designed, implemented, and verified the SLAM system that can be used on mobile devices. Mobile SLAM is composed of a stand-alone type that directly performs SLAM operation on a mobile device and a mapping server type that additionally configures a mapping server based on FastAPI to perform SLAM operation on the server and transmits data for map visualization to a mobile device. The mobile SLAM system proposed in this paper mixes the two types in order to make SLAM operation and map generation more efficient. The stand-alone type of SLAM system was configured as an Android app by porting the OpenVSLAM library to the Unity engine, and the map generation and performance were evaluated on desktop PCs and mobile devices. The mobile SLAM system in this paper is an open-source project, so it is expected to help develop AR contents based on SLAM in a mobile environment.

Background and objective: We measured the particulate matter (PM) reduction rate when the vegetation biofilter operated for 1 hour every 6 hours. As a result of data analysis, PM concentrations in the subway station showed a high correlation with the train stop duration depending on the service frequency of train operations, and the reduction rate was high in the section where PM concentrations as well as the change rates were high. Therefore, in order to maximize PM reduction using the biofilter, it is necessary to optimize the performance of the biofilter in various environments by designing an intelligent algorithm that takes the frequency of train operations into account.Methods: We collected PM10 and PM2.5 data for 2 weeks in position 5 (POS5) and POS6. The multi-module vegetation biofilter was installed in POS6, whereas it was not in POS5. The average hourly data for PM concentrations were obtained, and the change in concentrations was analyzed every 24 hours.Results: PM in subway stations is known to spread from the tracks to waiting areas as the train moves. Therefore, the change in PM concentrations by time zone has a high correlation of about 0.8 with the sum of the stop duration depending on the frequency of train operations. The vegetation biofilter is operated 4 times a day for 1 hour every 6 hours. The reduction rate reached up to 30% in 6-8 hours, where the PM concentrations increased sharply.Conclusion: We came up with an algorithm to maximize the PM reduction effect by identifying the effect using the vegetation biofilter as a means for air purification in public use facilities such as subway stations. To this end, there is a need for a method that can dynamically select the section where the PM concentrations increase rapidly and set the biofilter operation time.