Exploring Data Fusion for AI-Based People Counting in a Static Public Transport Vehicle Using Multiple IR-UWB Transceivers

People counting is a relevant task for many different applications in the sensor domain requiring real-time monitoring of a closed environment. One of these cases can be found in the public transport sector, where information such as the occupancy rate of each vehicle can be an important factor in decision making for related organizations as well as users. Impulse Radio Ultra-wideband (IR-UWB) sensors have proved to be a successful tool in such tasks, but challenges introduced by noise and multipath propagation of the signals can significantly hinder the performance of these systems. AI and Machine Learning have been widely utilized to address such challenges; however, inherent weaknesses related to data collection are an obstacle to the reliable deployment of these solutions in real-world scenarios. In this paper the feasibility of an AI-based UWB people counting system in a public transport vehicle has been studied. Four different experimental datasets have been produced in semi-static conditions, each being individually used to train and test four AI models to solve the task. Particular emphasis during the analysis was placed on the quantity and variety of data used for training, including the fusion of multiple sensor readings and the splitting of the datasets using two methods addressing different challenges. Results proved the fusion of data from different sources to be beneficial, as the models were able to correctly classify up to 100% of the signals from the test set when trained on an exhaustive amount of data. Performing inference on unseen static configurations has led to satisfactory results on simpler scenarios, with the best models attaining up to 91.5% accuracy, while achieving so under more difficult environments is left to subsequent research.