Privacy-centric AI-based real-time storage-less edge computing approaches for passenger counting and action classification on public transport vehicles

Abstract

Public transport operators aspire to provide the mobility solutions of the future. They attempt to ensure the safety and comfort of passengers in public transportation by analyzing their behavior. In this paper, we propose a novel approach for human action classification, falling detection, and passenger counting using a stream of 2D video frames. This could help the transport operators in planning transportation networks, improving the design of their vehicles, and enhancing passengers' flow throughout the day. Also, the falling detection could be used to ensure the passenger's safety onboard. This approach uses the body geometry and the rate of change in human body joints across consecutive frames to classify whether a person is either sitting or standing. In addition, the geometric model uses the previously predicted action for each person until the model detects a certain decrease or increase in the joints position of the human. Previously this was done by training models for classifying actions. These models were dependent on the environment they were used in and the dataset used for model training. On the other hand, upon comparing both our presented geometric approach and the classifier approach in terms of speed and accuracy, our approach superseded as there is no machine learning model used in the geometric approach other than the pose estimation model. Using our approach, we were able to develop a privacy-aware system that does not record anything, upload any user data to a server, or aggregate user information.