Kazumi Komiya

Accurate estimation of optical flow for fully automated tracking of moving-objects within video streams

This paper proposes a new approach to accurately estimate optical-flow for fully automated tracking of moving-objects within video streams. This approach consists of an occlusion-killer algorithm and template matching using segmented regions, which is performed by a skip-labeling algorithm. The skip-labeling algorithm can be used to segment an image into integrated regions of the same feature. The occlusion-killer algorithm is executed using the difference between the results of the forward and backward template matching in segmented frames. The occlusion-killer algorithm can be used to accurately estimate optical-flow in the neighboring regions of a moving-object on the background. A well-known optical-flow estimation algorithm for the block matching has problems of insufficient accuracy is limited due to block size, aperture, and occlusion. A new optical-flow estimation algorithm provides the solutions to these problems. Experiments are carried out to approve this new optical-flow estimation method can be applied to the objects of nonlinear movement.

Effective designation of specific shots on video service system utilizing Mahalanobis distance

A new video service system characterized by specific shot designation at high recall rate was built based on segmented motion vector magnitudes on each block of video frames, and inter-block correlations for making a decision logically. This video service system was applied to find out dramatic scenes from live video of baseball games.

Video service utilizing networked image retrieval system

Proposed is a new video service system utilizing the image directory and a flat-panel digital TV display, featuring the retrieval of a specific sequence of MPEG-2 video stream of a baseball game designated by the specific customer.

Video analysis based on motion vectors and its application to image database

Digitalized terrestrial broadcast can offer many programs to the TV viewer like communication satellite broadcasting. And it will spread in Japan. And image retrieval is necessary for us to watch many programs. Thereupon, we proposed method that particular scene is retrieved from video image. That retrieval method is to use the motion vector contained in MPEG-2 stream. Concretely, the movement of image is defined as the characteristics from information of motion vector. And the characteristic of a retrieval object and retrieval query scene are compared. In this way, target scene is retrieved. And we examined about efficacy of this retrieval method by using a television image.

A fully automated object extraction from video stream: a useful tool for distributed object-based browsing and content-based searching systems

This paper proposes an approach to automatically extract an accurate object from a video stream. The new approach consists of a skip-labeling algorithm for feature-based segmentation, an occlusion-killer algorithm for estimating accurately optical flow, and a shrink-merge tracking algorithm for tracking an object. The shrink-merge tracking algorithm is executed, based on the time continuity of moving-objects, using morphological image processing, such as dilation and erosion. The dilation and erosion are repeatedly executed using projection processing in which the object area in a following frame is derived from the object area in a current frame. The shrink-merge tracking algorithm can also project the area of a rotating object in a current frame on the rotating-object containing the newly appearing regions in the next frame. The automated object extraction method works satisfactory for objects which are moving nonlinearly within the video stream, and works satisfactory in 450 frames.

Dynamic load balancing for distributed movie based browser systems

This paper describes the dynamic load balancing of distributed movie based browser systems, which can visually be navigated by the integration of movies and linking-information. The linking-information indicates replicated multimedia information (such as sounds, videos and multiple super high definition (SHD) pictures) stored in each server on the network. Proposed new dynamic media partitioning performs the load-balancing of the respective server and network for multimedia information. The proposed method efficiently uses the servers and network, reserving and allocating the bandwidths of the transmission lines between a client and servers on the network, monitoring the load each server, and effectively partitioning continuous multimedia information. This method provides a solution to the problem caused by lack of the bandwidths that the server supplies to the client, or by lack of the bandwidths of the transmission lines between a client and a server on the network. The problem is derived from the deficiency of the throughput (64 Mbit/s) for lines with required high bandwidths (192 Mbit/s), long delay times (the round-trip time measures 8 ms) and a window size of 64 Kbytes. Improved dynamic media partitioning including a round-trip time parameter provides high throughput of 192 Mbits/sec by partitioning and allocating the load to 3 servers, giving a solution to this problem.

A study of color responses in visual evoked potentials using the kernel method

The relationship between color stimuli and VEPs was considered. VEPs elicited from normal subjects by chromatic modulated stimuli were measured, and their binary kernels were estimated by a nonlinear system identification method. Results showed that first-order binary kernels depended on the color of the stimulus. Opponent-color responses were considered to be included in first-order binary kernels.

Contact-type full-color reading system using three-color solid state illuminators

This paper reports a new contact-type, full-color reading system for compact color facsimiles. This system mainly consists of a direct-contact type image sensor and bright 3-color LED arrays. As the sensor does not require an optical lens system like a rod lens, no color aberrations occur, making this system suitable for color reading. Recent advances in brightness for blue LED have made a 3-color solid state document illuminator possible. An LED illuminator is superior to a fluorescent lamp in uniform illumination, compactness, and longevity. Specifically two points of system structure have been researched. One is an optical system design based on the light transfer performances. The second is a 3-color illuminating method which can reproduce color accurately with high speed scanning. The main experimental results are: (1) reading speed per line is 15 msec which is acceptable for G3 class color facsimile, and (2) resolution is 0.5 at 8 line/mm of spatial frequency for RGB color bar- chart, (3) color difference is 12 which is slightly higher than that of the ordinary systems but can be easily improved through signal processing techniques. Therefore, the proposed reading system is possible for application in compact color facsimiles.

Pattern edge color difference in a color image sensor with a GRIN rod lens.

The color distortion arising around pattern edges of pictures scanned by a color image sensor using a GRIN rod lens is simulated by a model that uses a Gaussian distribution function as the spread function of the lens for two color separation systems. In the time division system, the maximum permissible DeltaH(B) which indicates color aberration is obtained. Although this value depends on some sensor parameters, such as pixel size and pixel period, the method to obtain the maximum permissible DeltaH(B) presented in this paper is suitable in almost all cases. In the space division system, the effect of color aberration on color distortion is clarified, and it is seen that permissible color aberration is larger than that in the time division systems.

Optimum Number of Blocks on each frame dissected to perform video retrieval ---Pitching scene retrieval for baseball games in terms of human sensation ---

The precision rate was much lower than the recall rate, when high recall rate was desired, in video retrieval because of unwanted video scenes contained in the designated scenes set. This paper describes an experimental study on the test of how to obtain high precision rate keeping the recall rate high, visually confirming the optimum number of dissected blocks used to perform the video retrieval in each video frame of baseball games. The test was carried out to check if the subjects could recognize the desired scenes correctly when sets of contiguous video frames, each dissected into one of 2 by 2 through 20 by 20 blocks, were demonstrated to the respective subjects in a random order of frame dissection into blocks. The result of the test showed that a satisfactory score was obtained from the frame consisting of 10 by 10 blocks. The result implies that unwanted scenes will be rejected from the pitching scenes sets when 10 by 10 blocks will be chosen for the video retrieval.