Changseok Bae

A novel real time video tracking framework using adaptive discrete swarm optimization

A new adaptive discrete swarm optimization (ADSO) is proposed for video tracking.ADSO shows high accuracy rate and fast tracking and relocating speed.Error rate is reduced to 70.21% of Particle Swarm Optimization (PSO).Processing time per frame is reduced to 58.6% of PSO. This paper has proposed a new adaptive discrete swarm optimization (ADSO) for the video tracking framework. Each target object is first presented by a search window with four-dimensional features, which include 2D coordinates of the search window, its width and height. The image in the search window of a target object is extracted to calculate the HSV histograms, which are used to establish a feature model for the target object. Then the particles fly in a sub-search-space to find an optimal match of the target. If any occlusion or disappearance of the target object is detected, the particles will adaptively update their searching strategies in order to recapture the target. The experimental results demonstrate that the ADSO can out-perform the traditional PSO algorithm in the aspects of high accuracy rate and fast tracking and relocating speed.

A hybrid gravitational search algorithm with swarm intelligence and deep convolutional feature for object tracking optimization

Abstract Large number of object trackers based on particle swarm optimization (PSO) and its variants have been published in the recent decade. However, the majority of algorithms does not perform well when evaluated against the online object tracking benchmark. In the analysis of the existing swarm intelligence based object trackers, pre-mature convergence, loss in particle information and inadequate feature are identified as the factors that hinder the performance of this class of trackers. In this regard, this paper proposes to use the hybrid gravitational search algorithm (HGSA) to increase the utilization of particle information and to facilitate thorough search inside the video frame before convergence. HGSA elegantly combines GSAs gravitational update component with the cognitive and social components of PSO using a novel weight function. The hybridized algorithm acquires the exploitation of past information and fast convergence property typical of PSO, while retaining the GSA capability in fully utilizing all current information. Moreover, the incorporation of deep convolutional feature is proposed to address the inadequacy of the weak hue, saturation and value (HSV) histogram feature. Experimental results using videos from the online tracking benchmark show that the proposed HGSA tracker with deep convolutional feature (DeepHGSA) has increased accuracy of ADSO, the best existing Swarm Intelligence based tracker, by 50.6% and robustness by 56.9% measured by area under curve.

SDIF: Social device interaction framework for encounter and play in smart home service

In this paper, a framework for the mutual interaction of smart devices used in a smart home service is proposed by utilizing the concept of device sociality to minimize human interventions. The framework makes easy device collaboration possible by providing social relations among devices when they encounter each other. To measure the effect of the reduction in human intervention during the process of configuring the mutual interactions of such devices, a zero configuration index of the mutual interaction processes among smart devices was adopted. To show its feasibility in terms of minimizing human intervention, the proposed method was compared with another traditional device interaction protocol for a file and screen sharing application between two smart devices.

Error Detection and Copyrights Protection Scheme for MPEG-2 Based on Channel Coded Watermark Signal

This paper proposes an error detection and copyrights protection scheme for MPEG-2 video based on digital watermarking algorithm using channel coding. The watermark signal is generated by applying copyrights information of video data to convolutional encoder, and it is embedded into every block in intra frames while encoding to MPEG-2 video stream. In the decoder, the embedded signal is detected from each block of intra frames, and the detected signal is used to localize errors in the video stream. The detected signal can also be used to claim ownership of the video data by decoding it to copyrights information. In this stage, errors in the detected watermark signal can be corrected by channel decoder. Experimental results show that the proposed method can detect errors in the video stream while decoding and reconstruct copyrights information more correctly than the conventional method

Image Based Video Querying Algorithm Using 3-Level Haar Wavelet Transform Features

Surveillance cameras and smartphone cameras produce huge amount of video data in our daily life. Effective utilization of huge video data is emerging as a new big problem in the ubiquitous video intelligent systems. Searching for a specific frame in video stream is one of challenging issues in this area. This paper proposes an image based video querying algorithm using 3-level Haar wavelet transform features. Hierarchical decomposition of wavelet transform enables to use features in both space and scaling domains. This paper employs 3-level Haar wavelet feature of an image to query matched frame in a video stream. In experimental results, we can find that the proposed algorithm shows about 1–8% better performance in accuracy than other algorithms.

Estimation of Human Social Relation Based on Device Connectivity

People in these days use personal smart device as a typical device to communicate with others. Talks, texts, SNS, and even e-mails are available over the smartphones. Thus, we can consider the smartphones are able to represent personal closeness. This paper proposes an estimation model of human social relationship based on analysis of device connectivity. We have collected device-level communication data, and analyze them to find out correlation with human social relationships. Experimental results show that the proposed estimation model has linear relationship between device connectivity and human social relations. We can employ these results to construct sociality between devices.

Modeling of Verb Representation for Visual Data Understanding

This paper aims to propose a design and architecture of a lexicon of verbs that would help in parsing visual scenes into linguistically grounded representation as well as in generating the semantic representations. To this end, we look over verb representation in linguistics for a skeleton architecture of a model and observe visual data in order to extract detailed properties embedded in visual scenes. This paper mentions early results of modeling works so we need to explorer various factors and apply them in the modeling works continuously.

Biologically Inspired Computational Models of Visual Attention for Personalized Autonomous Agents: A Survey

Perception is one of essential capabilities for personalized autonomous agents that act like their users without intervention of the users in order to understand the environment for themselves like a human being. Visual perception in humans plays a major role to interact with objects or entities within the environment by interpreting their visual sensing information. The major technical obstacle of visual perception is to efficiently process enormous amount of visual stimuli in real-time. Therefore, computational models of visual attention that decide where to focus in the scene have been proposed to reduce the visual processing load by mimicking human visual system. This chapter provides the background knowledge of cognitive theories that the models were founded on and analyzes the computational models necessary to build a personalized autonomous agent that acts like a specific person as well as typical human beings.