Pedestrian tracking by using MPEG-based video signal processing (放送技術)

概要

論文の詳細を見る
In this paper, we propose a method for pedestrian detection and tracking which is based on MPEG video processing. Firstly AC power of frame difference for each block of pixels is calculated through DCT coefficients and is evaluated as a confidence measure of the pedestrian. Secondly, similarity of the motion between neighboring MPEG macroblocks is calculated and evaluated as a classification measure for moving objects. Finally these two measures are integrated and judged by the specified rule base to output recognition results. This first stage method has independently attained the maximum recognition rate of more than 90% per second. As the second stage algorithm for the moving camera, a multivariate time series analysis using the PCA (Principal Component Analysis) has been tried. Also future works on adaptation and behavior modeling are investigated.
社団法人映像情報メディア学会の論文
2008-12-11