4:50 PM - 5:10 PM
[3Q4-J-13-04] Likelihood distribution of Pedestrian Trajectories rendered by Variational Autoencoder
Keywords:probabilistic pedestrian behavior modeling, visualizing learned content
We studied applicability of Variational Autoencoder (VAE) to capture stochastic nature of pedestrian moves in a public space without explicit labels. Movies for training the network were recorded in a public pedestrian street and an exhibition booth. These movies were converted to grayscale images representing observed pedestrian locations and occupied areas. VAE was trained on 90% of data and rest of data was kept for validation. The validation result showed satisfactory reconstruction performance of pedestrian distributions in video frames. We propose a novel method to render our expectation of finding a pedestrian in a crowd as 2-D images by utilizing the trained network. Images rendered by this method correspond to subjective images usually only captured in our mind.