First-Person Vision Based Prediction of Information Seeking and Use Behavior in Library

Abstract

We propose an automated method for predicting subject behaviors based on first-person vision in an area surrounding bookshelves. The proposal classifies each frame within a movie recorded using a head-mounted camera to the six primitive behaviors according to naive Bayes nearest-neighbor method (NBNN). A prediction experiment is conducted using two image sequences recorded by a head-mounted camera. The experimental results confirm that the average classification rates for NBNN with random sampling (including principal components analysis) are improved from 0.09 to 0.13 for one data set and from 0.03 to 0.08 for the other data set compared with the bag-of-features and support vector machine combination results.