인파 밀집 분석을 위한 보행자 검출 방법에 대한 연구

임혜연(Hye-Youn Lim),안명수(MingShou An),강대성(Dae-Seong Kang) 한국정보기술학회 2024 한국정보기술학회논문지 Vol.22 No.3

Research on technology for detecting people in video has been steadily conducted, and various studies such as human tracking and behavior recognition for dense population analysis have recently been conducted. Pedestrian detection in dense crowds can cause problems such as poor accuracy and false detection due to cover or overlap between pedestrians. Therefore, this paper proposes an improvement method based on the YOLOv5 model to solve this problem. First, an attachment mechanism-based neural network was constructed in the network feature fusion stage to improve feature extraction performance and reduce the computational burden of non-maximum suppression(NMS). Next, we propose a structure to create and combine differential attention feature maps for overlapping targets by improving the resolution of pyramid features of each neural network layer through upsampling. Experiments on CrowdHuman datasets showed that the proposed method improved mAP 2.3% and loss 0.013 as a result of real-time detection tests.

출입 통제 시스템 개발을 위한 optical flow와 해마 신경망 알고리즘에 관한 연구

임혜연(Hye-Youn Lim),김대일(Dae-il Kim),강대성(Dae-Seong Kang) 한국정보기술학회 2008 한국정보기술학회논문지 Vol.6 No.6

In this paper, we propose following algorithm for the development of monitoring system about entrance limitation place. Information of moving object in video stream is analyzed using optical flow. Analyzed data is classified using PCA(Principal Component Analysis). After classification, only object information among data is restored and extracted. Object information abstracted is detected to only face area using map which is learned by DGTM(Dynamic Genetic Tree Map) about the feature of face area. Using hippocampal neural network which make use of the relation of interaction between hippocampal and cortex, information of entrance people is learned. After learning, through accord or discord with detection area, the permission of entrance is decided. Therefore, with this automatic system we can improve efficiency in the management aspect.

스테레오 카메라 기반의 코너점 알고리즘을 이용한 이동 객체 검출

임혜연(Hye-Youn Lim),강대성(Dae-Seong Kang) 한국정보기술학회 2012 Proceedings of KIIT Conference Vol.2012 No.11

본 논문에서는 스테레오 카메라 영상에서 코너점 알고리즘을 이용하여 이동 객체를 검출 및 추적하는 방법을 제안한다. 코너점 알고리즘 중 SIFT 알고리즘은 좋은 성능을 가지지만 처리 속도가 느리다는 단점이 있다. 따라서 SIFT의 속도를 개선하여 고속으로 코너점을 검출하는 SURF 알고리즘을 적용하였다. 객체 겹침일 때도 지속적으로 추적하기 위해 스테레오 카메라 기반으로 깊이 지도 영상을 이용하였으며 추적 알고리즘은 간단한 구현이 가능한 meanshift 알고리즘을 토대로 사용하였다. 그리고 고속으로 검출된 코너점들을 이용하여 움직이는 객체의 크기가 변화함에 따라 탐색 윈도우도 변화하도록 제안하였다. 즉, 조명이나 객체의 크기 변화에도 강인하게 추적할 수 있는 방법을 제안하였다. 실험을 통해 제안하는 방법이 기존의 방법보다 정확성에서 객체 추적이 효과적임을 증명하였다. In this paper, we suggest that a moving object detection and tracking using corner points algorithm based on a stereo camera. scale invariant feature transform (SIFT) algorithm has good processing, but slow processing time. Therefore, speeded up robust feature (SURF) algorithm which is improved to SIFT algorithm is applied to our method. Images based on a stereo camera used to the proposed method for continuous tracking during object occlusion and meanshift algorithm used tracking algorithm is able to simple implementation. We present robust object tracking method according to illumination and variable object size with generated adaptive search window using fast corner points detection of SURF algorithm. Comparative results from experiments show that the proposed algorithm has improved performances in terms of accuracy.

실외환경에서 효과적인 배경추출을 이용한 움직임 객체 추정

임혜연(Hye-Youn Lim),강대성(Dae-Seong Kang) 한국정보기술학회 2009 한국정보기술학회논문지 Vol.7 No.3

In this paper, we suggest an estimation technology of a moving object with a method of efficient background extraction in an outdoor environment with pixel based background subtraction technology. First, through the homomorphic filtering in preprocessing, sensitive actions for environment decrease. After preprocessing, background pixels are divided into static and dynamic pixel in a moving region. Background is extracted using dynamic background pixel. Therefore the processing time of each frame decreases. We make background updating algorithm using proper critical values. The processing time for each frame in the experimental results of proposed method is 10.394㎳, and precision is 89%. The experimental results show that the proposed method is more robust, accurate and powerful than existed methods.

객체 추적 시스템을 위한 컬러 확률 분포를 이용한 개선된 객체 추적 알고리즘에 관한 연구

임혜연(Hye-Youn Lim),강대성(Dae-Seong Kang) 한국정보기술학회 2010 한국정보기술학회논문지 Vol.8 No.2

In this paper, we present the moving objects tracking system by adaptive background modeling and objects' color probability distribution. First, to reduce noises of environment input sequences are preprocessed by Gaussian filtering. After preprocessing, the input sequences are divided into background and objects using adaptive Gaussian mixture model. To delete noises of objects, Detected objects are processed by morphology filter. Color based MeanShift or CAMShift algorithm must need designating initial search window. In this paper, Automatically assigning the search window with detected objects' location and region, our system can track the objects not designating initial search window The experimental results show that the proposed method is more fast, accurate and powerful than existed methods.

지식 증류를 이용한 DA-FSL 모델의 학습 수렴 곡선 및 학습 시간 개선 방법

임혜연(Hye-Youn Lim),이준목(Jun-Mock Lee),강대성(Dae-Seong Kang) 한국정보기술학회 2020 한국정보기술학회논문지 Vol.18 No.10

It is easy to train a machine learning model well in a field with a lot of data. However, it is not easy to train the learning model well where learning data is deficient. In order to solve this problem, knowledge distillation through transfer learning has recently attracted attention. Knowledge distillation refers to a methodology for transferring knowledge from a large number of large networks learned through ensemble techniques to one small network. In this paper, we propose a method to improve the learning accuracy as well as the learning time by applying the knowledge distillation technique to the DA-FSL model. Three knowledge distillation methods were used to apply it to the DA-FSL model, and through experiments, the curve converging to the initial target learning accuracy was evaluated. We compared the learning time required for model applying knowledge distillation through the proposed method and normal DA-FSL model without knowledge distillation. As a result, it was confirmed through an experiment that it was reduced from a minimum of 3% to a maximum of 35%.

고속의 강력한 특징점 기반의 향상된 이동 객체 검출 기법에 관한 연구

임혜연(Hye-Youn Lim),강대성(Dae-Seong Kang) 한국정보기술학회 2012 한국정보기술학회논문지 Vol.10 No.4

The advantage of meanshift or continuously adaptive mean shift (camshift) which is based on color probability distributions is robust about specific color objects, but the disadvantage of it is sensitive due to illumination and noise. Therefore, to offset this weakness, this paper is a study of enhanced moving object detection method based on speeded up robust feature points(SURF). The proposed method extracts the fast feature points with SURF algorithm of a detected object which is the region of interest (ROI), and generates a adaptive search window using the given information which is the positions of extracted feature points, and tracks moving objects through it. Comparative results from experiments show that the proposed algorithm is more 0.0126sec about processing time than existed methods. It will be used basic technology in a diversity of robot application fields via the moving object detection technique with the development of processing performance.

컬러 확률 분포를 이용한 움직이는 객체 추적 알고리즘에 관한 연구

임혜연(Hye-Youn Lim),강대성(Dae-Seong Kang) 제어로봇시스템학회 2009 제어로봇시스템학회 합동학술대회 논문집 Vol.2009 No.12

오늘날 여러 분야에서 지능형 감시카메라의 중요성이 부각되고 있다. 여러 기술의 밑바탕이 되는 객체를 검출하고 추적하는 기술의 성능을 높이기 위해 많은 연구가 진행 중이다. 본 논문에서 우리는 적응적인 배경 모델링을 생성하여 객체를 검출하고 검출된 객체의 색 정보를 이용하여 추적하는 방법을 제안한다. 잡음을 제거하기 위해 가우시안 필터링으로 전처리한 후, 배경을 모델링하여 가장 기본적이면서 중요한 기초 작업인 움직이는 객체와 배경을 분리한다. 이 때 적절한 파라미터들을 사용하여 적응적인 가우시안 혼합 모델링으로 움직이는 객체를 검출한다. CAMShift (Continuously Adaptive Mean Shift) 알고리즘을 사용하기 위해서는 추적할 객체의 위치와 크기를 먼저 지정해야 한다. 따라서 이 부분을 응용하여 배경 모델링으로 검출된 객체의 정보를 이용하여 지정하지 않고 자동으로 할당하는 개선된 CAMShift 알고리즘에 적용시켜 객체를 추적한다.

객체의 정확한 추적을 위한 특징점을 사용한 VSW 알고리즘에 관한 연구

임혜연(Hye-Youn Lim),강대성(Dae-Seong Kang) 한국정보기술학회 2011 한국정보기술학회논문지 Vol.9 No.1

Meanshift algorithm is object tracking according to color probability distributions. The advantage of this algorithm based color is robust about specific color objects but, the disadvantage of it is sensitive about non-specific color objects due to illumination and noise. Therefore, to supplement this weakness, in this paper, we present VSW(variable Search Window) algorithm which generated variable search window with robust feature points for the accurate tracking of moving objects. That is, the feature points of detected object(ROI:region of interest) are extracted using SIFT algorithm and then variable search window is generated by the information of them for the accurate tracking of moving objects. The experimental results show that the proposed method is more accurate and powerful than existed methods.

얼굴인식 시스템을 위한 특징 벡터 추출과 해마 신경망 알고리즘에 관한 연구

임혜연(Hye-Youn Lim),신정호(Jung-Ho Shin),강대성(Dae-Seong Kang) 한국정보기술학회 2008 Proceedings of KIIT Conference Vol.2008 No.-