연구단보: 얼굴 표정을 이용한 비접촉식 거짓말 탐지 시스템 개발 및 유의시점 분석

초록

In this study, facial expression changes were photographed and analyzed to develop a non-contact deception detection technique to solve the shortcomings of the existing contact deception detection. A simulated crime experiment was conducted to obtain video data for deception detection, and the data was collected using a POT (Peak of Tension Test). In order to analyze facial expression when lying, face landmarks were extracted from the face by using a model with four layers of stacked hourglass networks, a kind of CNN (Convolutional Neural Network). The deception detection was performed using stacked 66 face landmarks extracted in the frames from video in chronological order by the LSTM (Long Short Term Memory) model for truth/deception video classification, and the result using test set was 85.19%. LSTM activity analysis was performed to analyze the results of the model on which training was completed. Using the developed method, the possibility of deception detection using changes in facial expressions was confirmed, and important time for non-contact deception detection were identified.

키워드