【電機系書報討論演講訊息】6/1 鄭錫齊 教授：Techniques of Intelligent Computer Vision for Video Event Detection and Annotation

書報討論網址：http://www.ee.nthu.edu.tw/~ee591000/

演講公告網址：/p/403-1175-2629-1.php

題   目：Techniques of Intelligent Computer Vision for Video Event Detection and Annotation

講   員：鄭錫齊 教授 (海洋大學資工系)

時        間：2012年06月01日(星期五) 下午2時10分

地        點：資電館B01演講廳
 

Abstract:

This talk presents current trends on techniques of intelligent computer vision for video event detection and annotation. We first discuss a novel approach to incorporate temporal  information to generation a sequence of bag-of-words (BoW) models for video event recognition using string edit distance.  A traditional BoW model constructs feature vectors to be histograms of visual words without considering the temporal information regarding the arrangement of the visual words in the 2D image space. Our approach first segments the input video sequence into a set of video shots where each of them is further divided into multiple three dimensional video cubes. In this work we present a method to introduce temporal information within the BoW model by extracting space-time feature vectors from individual 3D cubes and learn the BoW codebook from these 3D cubes. Events are then modeled as a sequence of BoW feature vector. In addition, the system introduces temporal normalization to rearrange BoW feature vectors in a sequence for video event classification using a BoW features clustering approach.  A GHT-based event object detection and classification is also proposed to segment event objects from the input video clip. The BoW features extracted from these segmented event object can further enhance the classification accuracy of the system. Our framework presents a simple and effective way to infuse both temporal and spatial configurations.  Results show that the proposed method gives good performance on several publicly available datasets in terms of robustness and recognition rate.

歡  迎  聽  講