DEEPMATCH2: A comprehensive deep learning-based approach for in-vehicle presence detection

Abstract

The accurate detection of the mobile context information of public transportation vehicles and their passengers is a key feature to realize intelligent transportation systems. A topical example is in-vehicle presence detection that can, e.g., be used to ticket passengers automatically. Unfortunately, most existing solutions in this field suffer from low spatiotemporal accuracy which impedes their use in practice. In previous work, we addressed this challenge through a deep learning-based framework, called DeepMatch, that allows us to detect in-vehicle presence with a high degree of accuracy. DeepMatch utilizes the smartphone of a passenger to analyse and match the event streams of its own sensors with the event streams of counterpart sensors provided by a reference unit that is installed inside the vehicle. This is achieved through a new learning model architecture using Stacked Convolutional Autoencoders to compress sensor input streams by feature extraction and dimensionality reduction as well as a deep convolutional neural network to match the streams of the user phone and the reference device. The sensor stream compression is offloaded to the smartphone, while the matching is performed in a server. In this paper, we introduce DeepMatch2. It is an amended version of DeepMatch that reduces the amount of data to be transferred from the user and reference devices to the server by the factor of four. Further, DeepMatch2 improves the already good accuracy of DeepMatch from 97.81% to 98.51%. Moreover, we propose a travel inference algorithm, based on DeepMatch2, to detect the duration of whole passenger trips in public transport vehicles with a high degree of precision. This is needed to create intelligent and highly reliable auto-ticketing systems. Thanks to the high accuracy of 98.51% by DeepMatch2, the inferences can be carried out with a negligible error rate.