Detection of Texting While Walking in Occluded Scenarios Using Variational Autoencoder.

Texting while walking is a common behavior exhibited by pedestrians. While several studies explored the detection of texting while walking, the influence of occlusions were neglected. In this paper, we propose an image-based method which utilizes a pre-trained Variational Autoencoder. The proposed method takes sequence of 2D coordinates of upper body key points of the pedestrians as input, encodes the data into a 2D latent space, and uses the encoded data to distinguish text walkers from normal pedestrians. The proposed architecture enables the model to extract meaningful features from occluded data. Results of ablation test and comparison with a previous method revealed that the proposed architecture is successful in identifying text walkers even under heavy occlusion, outperforming a previously proposed method.