Introduction
[SS3-47] Head Orientation Estimation using Gait Observation
Gait Silhouette Volume
1gait
cycle
Estimation of ...
of 1

Head Orientation Estimation using Gait Observation (MIRU2014)

M. Nakazawa et al., "Head Orientation Estimation using Gait Observation," MIRU2014 You can download our paper from the following URL. https://www.jstage.jst.go.jp/article/ipsjtcva/6/0/6_63/_article M. Nakazawa, I. Mitsugami, H. Yamazoe and Y. Yagi, "Head Orientation Estimation using Gait Observation," IPSJ Transactions on Computer Vision and Applications, Vol.6, pp.63-67, 2014
Published on: Mar 3, 2016
Published in: Engineering      
Source: www.slideshare.net


Transcripts - Head Orientation Estimation using Gait Observation (MIRU2014)

  • 1. Introduction [SS3-47] Head Orientation Estimation using Gait Observation Gait Silhouette Volume 1gait cycle Estimation of Pedestrians’ head orientation is useful to estimate their intentions and behaviors  Not applicable for super low-resolution images where texture information is not rich enough Side Front Subjects were walking while paying attention to different directions (e.g. the walking direction, screens) • Silhouette images: captured from range data • Orientation ground truth: visually judged using RGB movies M. Nakazawa, I. Mitsugami, H. Yamazoe, Y. Yagi (Osaka Univ.) Gait feature space compressed by PCA 1 2 Overview of Our Method Benfold & Reid BMVC09  Unlike using facial region, gait feature of entire body is suitable for estimation from super low-res images Ex) Counting the number of pedestrians who watch signage carefully OU-ISIR Gait-Head Orientation Database Existing methods based on facial textures Our proposed method based on gait changes of entire body Gait feature calculation Head orientation discrimination Results Input images Gait Energy Image Comparison with other gait features Comp. with A non-linear classifier Diagonal Front Discriminant axes by LDA ClassA Front (#:224) ClassB D-Front (#:162) ClassC Side (#:77) Accuracy rate: leave-one-out cross validation (Average result of 50 iterations of 77 subjects randomly sampled from each class) 40 60 80 100 GEI FDF GEnI MGEI GFI CGI Accuracyrate[%] Front Diagonal front Side Kinects 40 60 80 100 Averageaccuracyrate[%] Only head Only torso Only legs Whole body Focusing on more dynamic gait info. GEI had the highest accuracy rates 40 60 80 100 LDA SVM Accuracyrate[%] Front D-front Side Contribution of body regions in lower-resolution cases Our method Averaged GSV in temporal axis Front D-front Side Front 84.9% 11.7% 3.5% D-front 8.4% 62.3% 29.3% Side 6.1% 36.8% 57.1% Comparison between two-class and three-class estimation Front D-front Side Gait change Large Small difference in the average Classifier selection was less important for this estimation  GEI captured sufficient features GEI Resolution (Pixel) Front Side Front 88.9% 11.1% Side 11.1% 88.9% Three-class estimation Two-class estimation Small The highest accuracy was reached in higher resolutions, but it became worse in lower resolutions Worse accuracy, but stable in even lower resolutions The highest accuracy among all in lower resolutions Our database contains 3 classes. It will be released soon!!!

Related Documents