• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.11a
• /
• pp.637-639
• /
• 2022

장면 전환 검출을 위해 히스토그램 등 단일 특성만 고려할 경우 움직임이 많은 영상과 조명의 변화가 다양하거나 단조로운 색상으로의 장면 전환이 일어나는 영상에 검출이 어렵다는 단점이 있다. 이를 개선하기 위해 프레임 간 색상 히스토그램을 비교 후, SURF를 적용하는 방법을 제안한다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.10
• /
• pp.4117-4135
• /
• 2020

Due to the angle of view, illumination and scene diversity, real-time dynamic face detection and recognition is no small difficulty in those unrestricted environments. In this study, we used the intrinsic correlation between detection and calibration, using a multi-task cascaded convolutional neural network(MTCNN) to improve the efficiency of face recognition, and the output of each core network is mapped in parallel to a compact Euclidean space, where distance represents the similarity of facial features, so that the target face can be identified as quickly as possible, without waiting for all network iteration calculations to complete the recognition results. And after the angle of the target face and the illumination change, the correlation between the recognition results can be well obtained. In the actual application scenario, we use a multi-camera real-time monitoring system to perform face matching and recognition using successive frames acquired from different angles. The effectiveness of the method was verified by several real-time monitoring experiments, and good results were obtained.

• International Journal of Fuzzy Logic and Intelligent Systems
• /
• v.11 no.2
• /
• pp.118-123
• /
• 2011

Video based object tracking normally deals with non-stationary image streams that change over time. Robust and real time moving object tracking is considered to be a problematic issue in computer vision. Multiple object tracking has many practical applications in scene analysis for automated surveillance. In this paper, we introduce a specified object tracking based particle filter used in an environment of multiple moving objects. A differential image region based tracking method for the detection of multiple moving objects is used. In order to ensure accurate object detection in an unconstrained environment, a background image update method is used. In addition, there exist problems in tracking a particular object through a video sequence, which cannot rely only on image processing techniques. For this, a probabilistic framework is used. Our proposed particle filter has been proved to be robust in dealing with nonlinear and non-Gaussian problems. The particle filter provides a robust object tracking framework under ambiguity conditions and greatly improves the estimation accuracy for complicated tracking problems.

• The Journal of Korea Robotics Society
• /
• v.18 no.3
• /
• pp.346-351
• /
• 2023

With the release of numerous open driving datasets, the demand for domain adaptation in perception tasks has increased, particularly when transferring knowledge from rich datasets to novel domains. However, it is difficult to solve the change 1) in the sensor domain caused by heterogeneous LiDAR sensors and 2) in the environmental domain caused by different environmental factors. We overcome domain differences in the semi-supervised setting with 3-stage model parameter training. First, we pre-train the model with the source dataset with object scaling based on statistics of the object size. Then we fine-tine the partially frozen model weights with copy-and-paste augmentation. The 3D points in the box labels are copied from one scene and pasted to the other scenes. Finally, we use the knowledge distillation method to update the student network with a moving average from the teacher network along with a self-training method with pseudo labels. Test-Time Augmentation with varying z values is employed to predict the final results. Our method achieved 3rd place in ECCV 2022 workshop on the 3D Perception for Autonomous Driving challenge.

• Science of Emotion and Sensibility
• /
• v.23 no.1
• /
• pp.17-28
• /
• 2020

Lowering observers' view-height may increase the amount of occlusion across objects in a visual scene and prevent the accurate identification of the objects in the scene. Based on this possibility, memory stimuli in relation to their expected views from different heights were displayed in this study. Thereafter, visual short-term memory (VSTM) performance for the stimuli was measured. In Experiment 1, the memory stimuli were presented on a grid-background drawn according to linear perspectives, which varied across observers' three different view-heights (high, middle, and low). This allowed the participants to remember both the color and position of each memory stimulus. The results revealed that testing participants' VSTM performance for the stimuli under a different memory load of two set-sizes (3 vs. 6) demonstrated an evident drop of performance in the lowest view-height condition. In Experiment 2, the performance for six stimuli with or without the grid-background was tested. A similar pattern of performance drop in the lowest condition as in Experiment 1 was found. These results indicated that different view-heights of an observer can change the amount of occlusion across objects in the visual field, and the sensory interference driven by the occlusion may further influence VSTM performance for those objects.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.44 no.3
• /
• pp.8-15
• /
• 2007

Recently, it is important to process multimedia data efficiently. Especially, in case of retrieval of multimedia information, technique of user interface and retrieval technique are necessary. This paper proposes a new technique which detects cuts effectively in compressed image information by MPEG. A cut is a turning point of scenes. The cut-detection is the basic work and the first-step for video indexing and retrieval. Existing methods have a weak point that they detect wrong cuts according to change of a screen such as fast motion of an object, movement of a camera and a flash. Because they compare between previous frame and present frame. The proposed technique detects shots at first using DC(Direct Current) coefficient of DCT(Discrete Cosine Transform). The database is composed of these detected shots. Features are extracted by HMMD color model and edge histogram descriptor(EHD) among the MPEG-7 visual descriptors. And detections are performed in sequence by the proposed matching technique. Through this experiments, an improved video segmentation system is implemented that it performs more quickly and precisely than existing techniques have.

• Journal of the Korea Society of Computer and Information
• /
• v.28 no.10
• /
• pp.85-92
• /
• 2023

In this paper, we propose a method for automatically generating story-based abbreviated summaries from long-form dramas and movies. From the shooting stage, the basic premise was to compose a frame with illusion of depth considering the golden division as well as focus on the object of interest to focus the viewer's attention in terms of content delivery. To consider how to extract the appropriate frames for this purpose, we utilized elemental techniques that have been utilized in previous work on scene and shot detection, as well as work on identifying focus-related blur. After converting the videos shared on YouTube to frame-by-frame, we divided them into a entire frame and three partial regions for feature extraction, and calculated the results of applying Laplacian operator and FFT to each region to choose the FFT with relative consistency and robustness. By comparing the calculated values for the entire frame with the calculated values for the three regions, the target frames were selected based on the condition that relatively sharp regions could be identified. Based on the selected results, the final frames were extracted by combining the results of an offline change point detection method to ensure the continuity of the frames within the shot, and an edit decision list was constructed to produce an abbreviated summary of 62.77% of the footage with F1-Score of 75.9%

• Journal of KIISE:Information Networking
• /
• v.34 no.3
• /
• pp.156-167
• /
• 2007

In this paper, we develop the model based prediction algorithm for Variable-Bit-Rate(VBR) video traffic with regular Group of Picture(GOP) pattern. We use multiplicative ARIMA process called GOP ARIMA (ARIMA for Group Of Pictures) as a base stochastic model. Kalman Filter based prediction algorithm consists of two process: GOP ARIMA modeling and prediction. In performance study, we produce three video traces (news, drama, sports) and we compare the accuracy of three different prediction schemes: Kalman Filter based prediction, linear prediction, and double exponential smoothing. The proposed prediction algorithm yields superior prediction accuracy than the other two. We also show that confidence interval analysis can effectively detect scene changes of the sample video sequence. The Kalman filter based prediction algorithm proposed in this work makes significant contributions to various aspects of network traffic engineering and resource allocation.

• The KIPS Transactions:PartD
• /
• v.11D no.1
• /
• pp.95-104
• /
• 2004

This paper proposes an efficient remote video evaluation system that is matched well with personalized characteristics of students using information filtering based on user profile. For making a question in forms of video, a key frame extraction method based on coordinate, size and color information is proposed. And Question-mating intervals are extracted using gray-level histogram difference and time window. Also, question-making method that combined category-based system with keyword-based system is used for efficient evaluation. Therefore, students can enhance their study achievement through both supplementing their inferior area and preserving their interest area.

• Proceedings of the Korean Information Science Society Conference
• /
• 2007.06c
• /
• pp.477-481
• /
• 2007

A background is a part where do not vary too much or frequently change in an image sequence. Using this assumption, it is presented a background acquisition algorithm for not only static but also dynamic view in this paper. For generating background, we detect a region, where has high correlation rate compared within selected region in the prior pyramid image, from the searching region in the current image. Between a detected region in the current image and a selected region in the prior image, we calculate movement vector for each regions in time sequence. After we calculate whole movement vectors for two successive images, vector histogram is used to determine the camera movement. The vector which has the highest density in the histogram is determined a camera movement. Using determined camera movement, we classify clusters based on pixel intensities which pixels are matched with prior pixels following camera movement. Finally we eliminate clusters which have lower weight than threshold, and combine remained clusters for each pixel to generate multiple background clusters. Experimental results show that we can automatically detect background whether camera move or not.