• Journal of the Korean Society of Industry Convergence
• /
• v.26 no.1
• /
• pp.113-119
• /
• 2023

In this paper, we propose an optimization of a pose estimation deep learning model for recognition of abnormal behavior in unmanned stores using radio frequencies. The radio frequency use millimeter wave in the 30 GHz to 300 GHz band. Due to the short wavelength and strong straightness, it is a frequency with less grayness and less interference due to radio absorption on the object. A millimeter wave radar is used to solve the problem of personal information infringement that may occur in conventional CCTV image-based pose estimation. Deep learning-based pose estimation models generally use convolution neural networks. The convolution neural network is a combination of convolution layers and pooling layers of different types, and there are many cases of convolution filter size, number, and convolution operations, and more cases of combining components. Therefore, it is difficult to find the structure and components of the optimal posture estimation model for input data. Compared with conventional millimeter wave-based posture estimation studies, it is possible to explore the structure and components of the optimal posture estimation model for input data using genetic algorithms, and the performance of optimizing the proposed posture estimation model is excellent. Data are collected for actual unmanned stores, and point cloud data and three-dimensional keypoint information of Kinect Azure are collected using millimeter wave radar for collapse and property damage occurring in unmanned stores. As a result of the experiment, it was confirmed that the error was moored compared to the conventional posture estimation model.

• Journal of Korea Multimedia Society
• /
• v.15 no.4
• /
• pp.508-515
• /
• 2012

This paper presents a method that estimates face pose for a given face image and synthesizes any posed face images using Active Appearance Model(AAM). The AAM that having been successfully applied to various applications is an example-based learning model and learns the variations of training examples. However, with a single model, it is difficult to handle large pose variations of face images. This paper proposes to build a model covering only a small range of angle for each pose. Then, with a proper model for a given face image, we can achieve accurate pose estimation and synthesis. In case of the model used for pose estimation was not trained with the angle to synthesize, we solve this problem by training the linear relationship between the models in advance. In the experiments on Yale B public face database, we present the accurate pose estimation and pose synthesis results. For our face database having large pose variations, we demonstrate successful frontal pose synthesis results.

• Journal of Astronomy and Space Sciences
• /
• v.41 no.2
• /
• pp.61-78
• /
• 2024

This study developed a real-time spacecraft pose estimation algorithm that combined a deep learning model and the least-squares method. Pose estimation in space is crucial for automatic rendezvous docking and inter-spacecraft communication. Owing to the difficulty in training deep learning models in space, we showed that actual experimental results could be predicted through software simulations on the ground. We integrated deep learning with nonlinear least squares (NLS) to predict the pose from a single spacecraft image in real time. We constructed a virtual environment capable of mass-producing synthetic images to train a deep learning model. This study proposed a method for training a deep learning model using pure synthetic images. Further, a visual-based real-time estimation system suitable for use in a flight testbed was constructed. Consequently, it was verified that the hardware experimental results could be predicted from software simulations with the same environment and relative distance. This study showed that a deep learning model trained using only synthetic images can be sufficiently applied to real images. Thus, this study proposed a real-time pose estimation software for automatic docking and demonstrated that the method constructed with only synthetic data was applicable in space.

• Journal of Broadcast Engineering
• /
• v.27 no.2
• /
• pp.174-184
• /
• 2022

3D human pose estimation models can be classified into a multi-view model and a single-view model. In general, the multi-view model shows superior pose estimation performance compared to the single-view model. In the case of the single-view model, the improvement of the 3D pose estimation performance requires a large amount of training data. However, it is not easy to obtain annotations for training 3D pose estimation models. To address this problem, we propose a method to generate pseudo ground-truths of multi-view human pose data from a multi-view model and exploit the resultant pseudo ground-truths to train a single-view model. In addition, we propose a multi-view consistency loss function that considers the consistency of poses estimated from multi-view images, showing that the proposed loss helps the effective training of single-view models. Experiments using Human3.6M and MPI-INF-3DHP datasets show that the proposed method is effective for training single-view 3D human pose estimation models.

• Journal of the Society of Disaster Information
• /
• v.20 no.2
• /
• pp.302-314
• /
• 2024

Purpose: This study proposes a fall detection model based on a top-down deep learning pose estimation model to automatically determine falls of multiple workers in an underground utility tunnel, and evaluates the performance of the proposed model. Method: A model is presented that combines fall discrimination rules with the results inferred from YOLOv8-pose, one of the top-down pose estimation models, and metrics of the model are evaluated for images of standing and falling two or fewer workers in the tunnel. The same process is also conducted for a bottom-up type of pose estimation model (OpenPose). In addition, due to dependency of the falling interference of the models on worker detection by YOLOv8-pose and OpenPose, metrics of the models for fall was not only investigated, but also for person. Result: For worker detection, both YOLOv8-pose and OpenPose models have F1-score of 0.88 and 0.71, respectively. However, for fall detection, the metrics were deteriorated to 0.71 and 0.23. The results of the OpenPose based model were due to partially detected worker body, and detected workers but fail to part them correctly. Conclusion: Use of top-down type of pose estimation models would be more effective way to detect fall of workers in the underground utility tunnel, with respect to joint recognition and partition between workers.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• fall
• /
• pp.187-188
• /
• 2021

본 논문에서는 Openpose의 신뢰도를 이용해 3D pose estimation의 정확도를 높이는 방법을 제안한다. 모델의 앞뒤양옆 네 방향에서 pose estimation의 진행하기 위해 3D 모델에 AABB(Axis Aligned Bound Box)를 생성한 다음, box의 네 옆면으로 모델을 투영시킨다. 각 면에 투사된 2D image에 대해 Openpose 2D pose estimation의 진행한다. 네 면에서 생성한 2D 스켈레톤들의 평균을 통해 3D 상의 교차점을 획득한다. Openpose에서 제공하는 신뢰도(confidence)를 이용하여 잘못 나온 2D 관절을 제외하는 것으로 더 정확한 pose estimation의 수행하였다. 실험적인 방법을 통해 신뢰도 0.45 이상의 값을 가지는 joint 만을 사용해 3D 교차점을 구함으로써 3D pose estimation의 정확도를 높였다.

• The Journal of Korea Robotics Society
• /
• v.12 no.1
• /
• pp.33-41
• /
• 2017

We present a region-based approach for accurate pose estimation of small mechanical components. Our algorithm consists of two key phases: Multi-view object co-segmentation and pose estimation. In the first phase, we explain an automatic method to extract binary masks of a target object captured from multiple viewpoints. For initialization, we assume the target object is bounded by the convex volume of interest defined by a few user inputs. The co-segmented target object shares the same geometric representation in space, and has distinctive color models from those of the backgrounds. In the second phase, we retrieve a 3D model instance with correct upright orientation, and estimate a relative pose of the object observed from images. Our energy function, combining region and boundary terms for the proposed measures, maximizes the overlapping regions and boundaries between the multi-view co-segmentations and projected masks of the reference model. Based on high-quality co-segmentations consistent across all different viewpoints, our final results are accurate model indices and pose parameters of the extracted object. We demonstrate the effectiveness of the proposed method using various examples.

• The Journal of Korea Robotics Society
• /
• v.11 no.4
• /
• pp.226-234
• /
• 2016

In this study, a model-referenced underwater navigation algorithm is proposed for high-precise underwater navigation using monocular vision near underwater structures. The main idea of this navigation algorithm is that a 3D model-based pose estimation is combined with the inertial navigation using an extended Kalman filter (EKF). The spatial information obtained from the navigation algorithm is utilized for enabling the underwater robot to navigate near underwater structures whose geometric models are known a priori. For investigating the performance of the proposed approach the model-referenced navigation algorithm was applied to an underwater robot and a set of experiments was carried out in a water tank.

• International journal of advanced smart convergence
• /
• v.7 no.2
• /
• pp.95-100
• /
• 2018

In this paper, we propose a binarized multi-scale module to accelerate the speed of the pose estimating deep neural network. Recently, deep learning is also used for fine-tuned tasks such as pose estimation. One of the best performing pose estimation methods is based on the usage of two neural networks where one computes the heat maps of the body parts and the other computes the part affinity fields between the body parts. However, the convolution filtering with a large kernel filter takes much time in this model. To accelerate the speed in this model, we propose to change the large kernel filters with binarized multi-scale modules. The large receptive field is captured by the multi-scale structure which also prevents the dropdown of the accuracy in the binarized module. The computation cost and number of parameters becomes small which results in increased speed performance.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.2 no.2
• /
• pp.120-133
• /
• 2008

This paper presents a novel approach for facial motion tracking and facial expression cloning to create a realistic facial animation of a 3D avatar. The exact head pose estimation and facial expression tracking are critical issues that must be solved when developing vision-based computer animation. In this paper, we deal with these two problems. The proposed approach consists of two phases: dynamic head pose estimation and facial expression cloning. The dynamic head pose estimation can robustly estimate a 3D head pose from input video images. Given an initial reference template of a face image and the corresponding 3D head pose, the full head motion is recovered by projecting a cylindrical head model onto the face image. It is possible to recover the head pose regardless of light variations and self-occlusion by updating the template dynamically. In the phase of synthesizing the facial expression, the variations of the major facial feature points of the face images are tracked by using optical flow and the variations are retargeted to the 3D face model. At the same time, we exploit the RBF (Radial Basis Function) to deform the local area of the face model around the major feature points. Consequently, facial expression synthesis is done by directly tracking the variations of the major feature points and indirectly estimating the variations of the regional feature points. From the experiments, we can prove that the proposed vision-based facial expression cloning method automatically estimates the 3D head pose and produces realistic 3D facial expressions in real time.