Segmentation of areas of interest inside a virtual reality store

Abstract

In consumer behavior studies, several signals like head position and eye-tracking, which are mostly unstructured, are recorded. Hence, the first step in these studies is to extract structured features. In feature extraction, segmenting the space into several Areas of Interests (AOI) can be beneficial. In this regard, these features are computed when the shopper is inside a specific zone or interacting with or looking at a specific area. One of the difficulties of this approach is defining AOIs. In this study, positional and eye-tracking data of 57 subjects were recorded in a virtual reality store using a Head Mounted Display. Each subject performed a free navigation task and the objective of the study was to classify the shoppers based on their genders. For this purpose, some AOI-based features were extracted from the behavioral data. The AOIs were cubic and defined with rectangles in zenithal perspective and the shelves levels in virtual store. Sizes of horizontal rectangles were then optimized using Genetic Algorithm (GA). In optimization, a cost function based on Fisher criterion is defined to maximize the linear separability between classes. After optimization, the features extracted with the optimized and several arbitrary AOIs are classified with Support Vector Machine method. The results show that gender classification accuracy with optimized AOIs is 85% and outperforms that of the other AOIs. Along with the outstanding results in this study, this methodology is capable of tuning other hyperparameters like navigational thresholds in classification problems.