• 대한임베디드공학회논문지
• /
• 제11권5호
• /
• pp.277-285
• /
• 2016

Moving object detection system has been an emerging research field in various advanced driver assistance systems (ADAS) and surveillance system. In this paper, we propose two optical flow based moving object detection methods at dynamic scenes. Both proposed methods consist of three successive steps; pre-processing, foreground segmentation, and post-processing steps. Two proposed methods have the same pre-processing and post-processing steps, but different foreground segmentation step. Pre-processing calculates mainly optical flow map of which each pixel has the amplitude of motion vector. Dense optical flows are estimated by using Farneback technique, and the amplitude of the motion normalized into the range from 0 to 255 is assigned to each pixel of optical flow map. In the foreground segmentation step, moving object and background are classified by using the optical flow map. Here, we proposed two algorithms. One is Gaussian mixture model (GMM) based background subtraction, which is applied on optical map. Another is adaptive thresholding based foreground segmentation, which classifies each pixel into object and background by updating threshold value column by column. Through the simulations, we show that both optical flow based methods can achieve good enough object detection performances in dynamic scenes.

• 정보처리학회논문지B
• /
• 제12B권7호
• /
• pp.721-728
• /
• 2005

본 논문에서는 영상 내 물체 영역에 대한 다중정규화와 움직임 색상 정보를 활용하여 이동 물체에 대한 후보 그룹을 추출하고 영상 분할 방법에 의해 대상 물체 영역을 정의하며 최종적으로 목표물체에 대한 검출방법을 제공하였다. 다중 색상변환에 의해 물체의 고유영역 확률을 강화하고 MCWUPC(Moving Color Weighted Unmatched Pixel Count) 연산을 활용하여 이동물체의 영역을 강조하는 두 가지 개념을 결합함으로써 최종적으로 입력 영상 시퀀스에서의 후보영역을 찾아 분할하였으며 매 프레임 정확한 물체의 외곽정보를 검출하였다. 제안된 알고리즘을 검증하기 위하여 이동물체의 이동 실시간이 가능한 시스템을 구축하였고, 다양한 배경을 포함한 실험영상 120 프레임을 처리한 결과 $89\%$ 이상의 추적 성공률을 보여주었다.

• 방송공학회논문지
• /
• 제25권7호
• /
• pp.1081-1094
• /
• 2020

With the recent development of deep learning, most computer vision-related tasks are being solved with deep learning-based network technologies such as CNN and RNN. Computer vision tasks such as object detection or object segmentation use intermediate features extracted from the same backbone such as Resnet or FPN for training and inference for object detection and segmentation. In this paper, an experiment was conducted to find out the compression efficiency and the effect of encoding on task inference performance when the features extracted in the intermediate stage of CNN are encoded. The feature map that combines the features of 256 channels into one image and the original image were encoded in HEVC to compare and analyze the inference performance for object detection and segmentation. Since the intermediate feature map encodes the five levels of feature maps (P2 to P6), the image size and resolution are increased compared to the original image. However, when the degree of compression is weakened, the use of feature maps yields similar or better inference results to the inference performance of the original image.

• 로봇학회논문지
• /
• 제17권3호
• /
• pp.255-263
• /
• 2022

3D object detection pipelines often incorporate RGB-based object detection methods such as YOLO, which detects the object classes and bounding boxes from the RGB image. However, in complex environments where objects are heavily cluttered, bounding box approaches may show degraded performance due to the overlapping bounding boxes. Mask based methods such as Mask R-CNN can handle such situation better thanks to their detailed object masks, but they require much longer time for data preparation compared to bounding box-based approaches. In this paper, we present a 3D object recognition pipeline which uses either the YOLO or Mask R-CNN real-time object detection algorithm, K-nearest clustering algorithm, mask reduction algorithm and finally Principal Component Analysis (PCA) alg orithm to efficiently detect 3D poses of objects in a complex environment. Furthermore, we also present an improved YOLO based 3D object detection algorithm that uses a prioritized heightmap clustering algorithm to handle overlapping bounding boxes. The suggested algorithms have successfully been used at the Artificial-Intelligence Robot Challenge (ARC) 2021 competition with excellent results.

• ETRI Journal
• /
• 제46권2호
• /
• pp.307-322
• /
• 2024

Most object detection methods use a horizontal bounding box that causes problems between adjacent objects with arbitrary directions, resulting in misaligned detection. Hence, the horizontal anchor should be replaced by a rotating anchor to determine oriented bounding boxes. A two-stage process of delineating a horizontal bounding box and then converting it into an oriented bounding box is inefficient. To improve detection, a box-boundary-aware vector can be estimated based on a convolutional neural network. Specifically, we propose a ResNeXt101 encoder to overcome the weaknesses of the conventional ResNet, which is less effective as the network depth and complexity increase. Owing to the cardinality of using a homogeneous design and multi-branch architecture with few hyperparameters, ResNeXt captures better information than ResNet. Experimental results demonstrate more accurate and faster oriented object detection of our proposal compared with a baseline, achieving a mean average precision of 89.41% and inference rate of 23.67 fps.

• 센서학회지
• /
• 제28권3호
• /
• pp.157-163
• /
• 2019

The LiDAR sensor, which is widely regarded as one of the most important sensors, has recently undergone active commercialization owing to the significant growth in the production of ADAS and autonomous vehicle components. The LiDAR sensor technology involves radiating a laser beam at a particular angle and acquiring a three-dimensional image by measuring the lapsed time of the laser beam that has returned after being reflected. The LiDAR sensor has been incorporated and utilized in various devices such as drones and robots. This study focuses on object detection and recognition by employing sensor fusion. Object detection and recognition can be executed as a single function by incorporating sensors capable of recognition, such as image sensors, optical sensors, and propagation sensors. However, a single sensor has limitations with respect to object detection and recognition, and such limitations can be overcome by employing multiple sensors. In this paper, the performance of an eight-channel scanning LiDAR was evaluated and an object detection algorithm based on it was implemented. Furthermore, object detection characteristics during daytime and nighttime in a real road environment were verified. Obtained experimental results corroborate that an excellent detection performance of 92.87% can be achieved.

• 한국컴퓨터정보학회논문지
• /
• 제12권2호
• /
• pp.71-76
• /
• 2007

주어진 영상 내에서 의미 있는 객체의 존재여부를 판단하는 객체검출과정은 컴퓨터 비전 시스템을 이용한 영상인식의 핵심요소 가운데 하나이다. 템플릿 정합 방식의 객체 검출방식에서 사용되어 얼굴 인식 등에서 좋은 성능을 보이는 Hausdorff 거리는 점의 집합으로 표현된 두 객체 사이의 다른 정도를 정의함으로써 주어진 모델과 가장 차이가 적은 객체를 검출하도록 한다. 본 논문에서는 일반적으로 사용되는 바와 같이 점의 집합이 영상의 에지를 검출하여 얻어진 경우, 에지가 가지고 있는 방향정보를 이용함으로써 검출의 정확성을 높일 수 있는 방법을 제안하고 있다. 일반적으로 사용되는 Euclidean 거리 대신에 방향을 가진 두 점을 일치시키는데 필요한 탄성에너지를 정의함으로써 검출의 정확성을 높이는 방법을 제시하고 있다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제14권12호
• /
• pp.4795-4815
• /
• 2020

Drone detection can be considered as a specific sort of small object detection, which has always been a challenge because of its small size and few features. For improving the detection rate of drones, we design a Deeper SSD network, which uses large-scale input image and deeper convolutional network to obtain more features that benefit small object classification. At the same time, in order to improve object classification performance, we implemented the up-sampling modules to increase the number of features for the low-level feature map. In addition, in order to improve object location performance, we adopted the down-sampling modules so that the context information can be used by the high-level feature map directly. Our proposed Deeper SSD and its variants are successfully applied to the self-designed drone datasets. Our experiments demonstrate the effectiveness of the Deeper SSD and its variants, which are useful to small drone's detection and recognition. These proposed methods can also detect small and large objects simultaneously.

• 디지털산업정보학회논문지
• /
• 제4권2호
• /
• pp.15-19
• /
• 2008

In this paper, we propose a new algorithm of image depth detection using motion estimation and object tracking. In industry, robots are used for automobile, conveyer system, etc. But, these have much necessary time. Thus, in this paper, we develop the efficient method of image depth detection based on motion estimation and object tracking.

• 한국정보기술학회논문지
• /
• 제17권5호
• /
• pp.13-20
• /
• 2019

최근 CNN을 기반으로 한 객체 검출 기술의 연구가 활발하다. 객체 검출 기술은 자율주행차, 지능형 영상분석 등에서 중요한 기술로 사용된다. 본 논문에서는 CNN 기반의 객체 검출기 중 하나인 SSD(Single Shot Multibox Detector)에 MI-FL(Moment Invariant-Feature Layer)을 적용하여 회전 변형에 강인한 객체 검출 시스템을 제안한다. 먼저 VGG 네트워크를 기반으로 입력 이미지의 특징을 추출한다. 그 후 총 6개의 특징 계층(Feature layer)을 적용하여 객체의 위치 정보와 종류를 예측해 경계 박스들을 생성한다. 그 후 NMS 알고리즘을 이용해 가장 객체일 확률이 높은 경계 박스를 얻는다. 하나의 객체 경계 박스가 정해지면 MI-FL을 이용해 해당 영역의 불변 모멘트 특징을 추출하여 미리 저장하고 학습한다. 이후 검출 과정에서 미리 저장해둔 불면모멘트 특징 정보를 이용해 검출함으로써 회전된 이미지에 대해 기존 방법보다 더 강인한 검출이 가능하다. 기존의 SSD와 MI-FL을 적용한 SSD의 비교를 통해 약 4~5%의 성능 향상을 확인하였다.