• 융합신호처리학회논문지
• /
• 제22권1호
• /
• pp.30-37
• /
• 2021

카메라 영상을 이용한 3차원 객체 추적 기술은 증강현실 응용 분야를 위한 핵심 기술이다. 영상 분류, 객체 검출, 영상 분할과 같은 컴퓨터 비전 작업에서 CNN(Convolutional Neural Network)의 인상적인 성공에 자극 받아, 3D 객체 추적을 위한 최근의 연구는 딥러닝(deep learning)을 활용하는 데 초점을 맞추고 있다. 본 논문은 이러한 딥러닝을 활용한 3차원 객체 추적 방법들을 살펴본다. 딥러닝을 활용한 3차원 객체 추적을 위한 주요 방법들을 설명하고, 향후 연구 방향에 대해 논의한다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제13권3호
• /
• pp.1450-1463
• /
• 2019

The problem towards crowd behavior recognition in a serious clustered scene is extremely challenged on account of variable scales with non-uniformity. This paper aims to propose a crowed behavior classification framework based on a transferring hybrid network blending 3D res-net with inception-v3. First, the 3D res-inception network is presented so as to learn the augmented visual feature of UCF 101. Then the target dataset is applied to fine-tune the network parameters in an attempt to classify the behavior of densely crowded scenes. Finally, a transferred entropy function is used to calculate the probability of multiple labels in accordance with these features. Experimental results show that the proposed method could greatly improve the accuracy of crowd behavior recognition and enhance the accuracy of multiple label classification.

• 방송공학회논문지
• /
• 제23권5호
• /
• pp.614-621
• /
• 2018

본 논문에서는 generative adversarial network (GAN)을 이용한 비감독 학습을 통해 깊이 카메라로 깊이 영상을 취득할 때 발생한 손실된 부분을 복원하는 기법을 제안한다. 제안하는 기법은 3D morphable model convolutional neural network (3DMM CNN)와 large-scale CelebFaces Attribute (CelebA) 데이터 셋 그리고 FaceWarehouse 데이터 셋을 이용하여 학습용 얼굴 깊이 영상을 생성하고 deep convolutional GAN (DCGAN)의 생성자(generator)와 Wasserstein distance를 손실함수로 적용한 구별자(discriminator)를 미니맥스 게임기법을 통해 학습시킨다. 이후 학습된 생성자와 손실 부분을 복원해주기 위한 새로운 손실함수를 이용하여 또 다른 학습을 통해 최종적으로 깊이 카메라로 취득된 얼굴 깊이 영상의 손실 부분을 복원한다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제14권3호
• /
• pp.1121-1141
• /
• 2020

We propose a robust visual object tracking algorithm fusing a convolutional neural network tracker trained offline from a large number of video repositories and a color histogram based tracker to track objects for mixing immersive audio. Our algorithm addresses the problem of occlusion and large movements of the CNN based GOTURN generic object tracker. The key idea is the offline training of a binary classifier with the color histogram similarity values estimated via both trackers used in this method to opt appropriate tracker for target tracking and update both trackers with the predicted bounding box position of the target to continue tracking. Furthermore, a histogram similarity constraint is applied before updating the trackers to maximize the tracking accuracy. Finally, we compute the depth(z) of the target object by one of the prominent unsupervised monocular depth estimation algorithms to ensure the necessary 3D position of the tracked object to mix the immersive audio into that object. Our proposed algorithm demonstrates about 2% improved accuracy over the outperforming GOTURN algorithm in the existing VOT2014 tracking benchmark. Additionally, our tracker also works well to track multiple objects utilizing the concept of single object tracker but no demonstrations on any MOT benchmark.

• 한국인터넷방송통신학회논문지
• /
• 제20권2호
• /
• pp.149-155
• /
• 2020

한국도로교통공단은 교통사고분석시스템(TAAS)을 활용하여 2015년부터 발생한 교통사고 원인을 분석한 통계를 제공하고 있다. 교통사고 발생 주요 원인으로, 2018년 한해 전체 교통사고 발생원인 중 전방주시 부주의가 대부분의 원인임을 TAAS를 통해 발표했다. 교통사고 원인에 대한 통계자료의 세부항목으로 운전 중 스마트폰 사용, DMB 시청 등의 안전운전 불이행 51.2%와 안전거리 미확보 14%, 보행자 보호의무 위반 3.6% 등으로, 전체적으로 68.8%의 비율을 보여준다. 본 논문에서는 Deep Learning의 알고리듬 중 CNN(Convolutional Neural Network)를 활용하여 첨단 운전자 보조 시스템 ADAS(Advanced Driver Assistance Systems)을 개선한 시스템을 제안하고자 한다. 제안된 시스템은 영상처리에 주로 사용되는 Conv2D 기법을 사용하여 운전자의 얼굴과 눈동자의 조향을 분류하는 모델을 학습하고, 차량 전방에 부착된 카메라로 자동차의 주변 object를 인지 및 검출하여 주행환경을 인지한다. 그 후, 학습된 시선 조향모델과 주행환경 데이터를 사용하여 운전자의 시선과 주행환경에 따라, 위험요소를 3단계로 분류하고 검출하여 운전자의 전방 및 사각지대 보조한다.

• 한국안전학회지
• /
• 제36권1호
• /
• pp.18-25
• /
• 2021

In many developed countries, such as South Korea, efficiently maintaining the aging infrastructures is an important issue. Currently, inspectors visually inspect the infrastructure for maintenance needs, but this method is inefficient due to its high costs, long logistic times, and hazards to the inspectors. Thus, in this paper, a novel crack inspection approach for concrete bridges is proposed using integrated image processing techniques. The proposed approach consists of four steps: (1) training a deep learning model to automatically detect cracks on concrete bridges, (2) acquiring in-situ images using a drone, (3) generating orthomosaic images based on 3D modeling, and (4) detecting cracks on the orthmosaic image using the trained deep learning model. Cascade Mask R-CNN, a state-of-the-art instance segmentation deep learning model, was trained with 3235 crack images that included 2415 hard negative images. We selected the Tancheon overpass, located in Seoul, South Korea, as a testbed for the proposed approach, and we captured images of pier 34-37 and slab 34-36 using a commercial drone. Agisoft Metashape was utilized as a 3D model generation program to generate an orthomosaic of the captured images. We applied the proposed approach to four orthomosaic images that displayed the front, back, left, and right sides of pier 37. Using pixel-level precision referencing visual inspection of the captured images, we evaluated the trained Cascade Mask R-CNN's crack detection performance. At the coping of the front side of pier 37, the model obtained its best precision: 94.34%. It achieved an average precision of 72.93% for the orthomosaics of the four sides of the pier. The test results show that this proposed approach for crack detection can be a suitable alternative to the conventional visual inspection method.

• Geomechanics and Engineering
• /
• 제37권1호
• /
• pp.49-64
• /
• 2024

Rocks undergoing repeated loading and unloading over an extended period, such as due to earthquakes, human excavation, and blasting, may result in the gradual accumulation of stress and deformation within the rock mass, eventually reaching an unstable state. In this study, a CNN-CCM is proposed to address the mechanical behavior. The structure and hyperparameters of CNN-CCM include Conv2D layers × 5; Max pooling2D layers × 4; Dense layers × 4; learning rate=0.001; Epoch=50; Batch size=64; Dropout=0.5. Training and validation data for deep learning include 71 rock samples and 122,152 data points. The AI Rock Constitutive Model learned by CNN-CCM can predict strain values(ε₁) using Mass (M), Axial stress (σ₁), Density (ρ), Cyclic number (N), Confining pressure (σ₃), and Young's modulus (E). Five evaluation indicators R², MAPE, RMSE, MSE, and MAE yield respective values of 0.929, 16.44%, 0.954, 0.913, and 0.542, illustrating good predictive performance and generalization ability of model. Finally, interpreting the AI Rock Constitutive Model using the SHAP explaining method reveals that feature importance follows the order N > M > σ₁ > E > ρ > σ₃.Positive SHAP values indicate positive effects on predicting strain ε₁ for N, M, σ₁, and σ₃, while negative SHAP values have negative effects. For E, a positive value has a negative effect on predicting strain ε₁, consistent with the influence patterns of conventional physical rock constitutive equations. The present study offers a novel approach to the investigation of the mechanical constitutive model of rocks under cyclic loading and unloading conditions.

• 정보처리학회논문지:컴퓨터 및 통신 시스템
• /
• 제11권3호
• /
• pp.99-104
• /
• 2022

대부분 첨단 동작 인식 컨볼루션 네트워크는 RGB 스트림과 광학 흐름 스트림, 양 스트림 아키텍처를 기반으로 하고 있다. RGB 프레임 스트림은 모양 특성을 나타내고 광학 흐름 스트림은 동작 특성을 해석한다. 그러나 광학 흐름은 계산 비용이 매우 높기 때문에 동작 인식 시간에 지연을 초래한다. 이에 양 스트림 네트워크와 교사-학생 아키텍처에서 영감을 받아 행동 인식을 위한 새로운 네트워크 디자인을 개발하였다. 제안 신경망은 두 개의 하위 네트워크로 구성되어있다. 즉, 교사 역할을 하는 광학 흐름 하위 네트워크와 학생 역할을 하는 RGB 프레임 하위 네트워크를 연결하였다. 훈련 단계에서 광학 흐름의 특징을 추출하고 교사 서브 네트워크를 훈련시킨 다음 그 특징을 학생 서브 네트워크를 훈련시키기 위한 기준선으로 지정하여 학생 서브 네트워크에 전송한다. 테스트 단계에서는 광학 흐름을 계산하지 않고 대기 시간이 줄어들도록 학생 네트워크만 사용한다. 제안 네트워크는 실험을 통하여 정확도 면에서 일반 이중 스트림 아키텍처에 비해 높은 정확도를 보여주는 것을 확인하였다.

• International journal of advanced smart convergence
• /
• 제10권3호
• /
• pp.163-171
• /
• 2021

Recently, Machine Learning-based visualization approaches have been proposed to combat the problem of malware detection. Unfortunately, these techniques are exposed to Adversarial examples. Adversarial examples are noises which can deceive the deep learning based malware detection network such that the malware becomes unrecognizable. To address the shortcomings of these approaches, we present Block-matching and 3D filtering (BM3D) algorithm and deep image prior based denoising technique to defend against adversarial examples on visualization-based malware detection systems. The BM3D based denoising method eliminates most of the adversarial noise. After that the deep image prior based denoising removes the remaining subtle noise. Experimental results on the MS BIG malware dataset and benign samples show that the proposed denoising based defense recovers the performance of the adversarial attacked CNN model for malware detection to some extent.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제12권11호
• /
• pp.5555-5567
• /
• 2018

State-of-art instance segmentation networks are successful at generating 2D segmentation mask for region proposals with highest classification score, yet 3D object segmentation task is limited to geocentric embedding or detector of Sliding Shapes. To this end, we propose an amodal 3D instance segmentation network called A3IS-CNN, which extends the detector of Deep Sliding Shapes to amodal 3D instance segmentation by adding a new branch of 3D ConvNet called A3IS-branch. The A3IS-branch which takes 3D amodal ROI as input and 3D semantic instances as output is a fully convolution network(FCN) sharing convolutional layers with existing 3d RPN which takes 3D scene as input and 3D amodal proposals as output. For two branches share computation with each other, our 3D instance segmentation network adds only a small overhead of 0.25 fps to Deep Sliding Shapes, trading off accurate detection and point-to-point segmentation of instances. Experiments show that our 3D instance segmentation network achieves at least 10% to 50% improvement over the state-of-art network in running time, and outperforms the state-of-art 3D detectors by at least 16.1 AP.