항공 라이다 점군 분류를 위한 기하학적 특성정보 기반 딥러닝 (Geometric Feature Information Based Deep Learning for Airborne LiDAR Point Cloud Classification)

서지정보

· 발행기관 : 한국측량학회
· 수록지 정보 : 한국측량학회지 / 42권 / 5호 / 395 ~ 409페이지
· 저자명 : 형성웅, 김경휘, 임재형, 최현석, 이동천

초록

사진측량은 다양한 플랫폼에 탑재할 수 있는 여러 종류의 첨단 센서와 사용자 중심의 성과물들을 효율적으로 생성할 수 있는 딥러닝(DL)과 같은 기술 발전으로 많은 분야에서 활용이 급증하고 있다. DL의 성능은 인공신경망 구조, 학습 데이터 및 학습방법 등 다양한 요소에 영향을 받는다. 본 논문은 라이다 (LiDAR: Light Detection and Ranging) 데이터의 3D 좌표에서 생성한 기하학적 특징을 활용하여 라이다 점군 분류를 위한 다중 모드 DL을 제시한다. 점군의 고유값으로부터 계산한 기하학적 특성인 omnivariance, eigenentropy, anisotropy, surface variation, sphericity, verticality를 학습에 활용하였다. 각각의 특성은 데이터에 내재된 객체들의 고유 정보이며, 이를 DL 모델 학습에 사용하여 시너지 효과에 의한 DL 성능을 향상시킬 수 있다. 또한, 다중 모드 DL에서 융합이 중요하며, 일반적으로 초기융합, 중간융합 그리고 후기융합으로 분류된다. 본 논문에서는 초기융합과 후기융합 기반의 혼합(또는 하이브리드) 방법을 사용하여 라이다 점군의 분류 결과를 분석하였다. 라이다 데이터와 기하학적 특징들을 멀티 모드 DL 학습에 사용하고 초기융합 방법을 수행하여 라이다 데이터의 분류 정확도가 최대 35% 향상되었다. 그러므로 다중 모드 DL은 데이터에 내재된 특성정보를 활용하는 효과적인 학습방법이다.

영어초록

The use of photogrammetry is rapidly increasing in many fields due to the development of various types of advanced sensors that can be mounted on a variety of payloads, along with state-of-the-art data processing technology, including DL (Deep Learning), which can efficiently produce user-oriented spatial information products. The performance of DL is affected by various factors, including neural network architecture, training data and method. This paper presents multimodal DL for LiDAR (Light Detection and Ranging) point cloud classification by utilizing geometric features derived from the 3D coordinates of LiDAR data. In particular, omnivariance, eigenentropy, anisotropy, surface variation, sphericity, and verticality, as geometric features computed from the eigenvalues of the point clouds, were utilized for training the DL model. Each feature represents unique intrinsic information about the objects. By revealing these characteristics inherent in the 3D coordinates of LiDAR data, a synergy effect in DL model training can be achieved to improve DL performance. Additionally, fusion is an important issue in multimodal DL. In this paper, we analyzed the classification from early-fusion and hybrid method based on late-fusion. The overall accuracy of the classification improved by up to 35% for test data by utilizing geometric features with early-fusion. Therefore, multimodal DL could be an effective training strategy by utilizing intrinsic feature information. 'The use of photogrammetry is rapidly increasing in many fields due to the development of various types of advanced sensors that can be mounted on a variety of payloads, along with state-of-the-art data processing technology, including DL (Deep Learning), which can efficiently produce user-oriented spatial information products. The performance of DL is affected by various factors, including neural network architecture, training data and method. This paper presents multimodal DL for LiDAR (Light Detection and Ranging) point cloud classification by utilizing geometric features derived from the 3D coordinates of LiDAR data. In particular, omnivariance, eigenentropy, anisotropy, surface variation, sphericity, and verticality, as geometric features computed from the eigenvalues of the point clouds, were utilized for training the DL model. Each feature represents unique intrinsic information about the objects. By revealing these characteristics inherent in the 3D coordinates of LiDAR data, a synergy effect in DL model training can be achieved to improve DL performance. Additionally, fusion is an important issue in multimodal DL. In this paper, we analyzed the classification from early-fusion and hybrid method based on late-fusion. The overall accuracy of the classification improved by up to 35% for test data by utilizing geometric features with early-fusion. Therefore, multimodal DL could be an effective training strategy by utilizing intrinsic feature information.

참고자료

· 없음