Abstract Human Activity Recognition (HAR) allows for unobtrusive indoor monitoring, particularly in elderly care. However, existing HAR methods face significant challenges due to the variability in home layouts, sensor types, and activity labels across different datasets, which limits their generalization and scalability. Most approaches require extensive customization, making cross-environment HAR implementation challenging in real-world scenarios. To address these challenges, we propose a unified HAR framework that introduces Functional Areas, which abstract physical spaces into standardized activity zones, and Detector Units, which map heterogeneous sensor configurations into a common representation. We evaluate our framework using multiple publicly available HAR datasets based on ambient sensor data of smart homes, testing two model architectures: a Holistic Approach, which trains a single GRU-based neural network on the combined datasets, and a Reductionist Approach, which employs an ensemble bagging method. The Holistic Approach demonstrated superior generalisation, achieving 0.84 precision and 0.73 accuracy, outperforming the reductionist approach.

Abstract This paper deals with the Bayesian safety assessment of autonomous vehicles (AV) using as a key safety measure the probability of catastrophic failure per mile of driving (pfm), assumed a random variable. The paper takes the view that pfm may (and typically will) vary due to changing road driving conditions. Accommodating this variation in a Bayesian inference on pfm requires one to use a multivariate probabilistic model whereby the changeable pfm is captured explicitly for the different driving conditions. The model that we use in this work is derived from our prior work and accounts for the uncertainties in both—the operational profile (i.e., the likelihood of the different driving conditions) and the pfms, conditional on the respective operating conditions. The concept of the “dynamic AV safety assessment (DyAVSA)” is presented in the work, too, whereby the Bayesian predictions used at run time rely on the operational data collected by a fleet of AVs. DyAVSA benefits both: (1) the AV vendors, for monitoring the safety changes of the entire AV fleet; (2) the owners/users of individual AVs, whose safety assessment is personalized and different from the assessment of the AV fleet. DyAVSA thus offers a major change in the AV safety management than is currently the case. It allows the AV users/owners to benefit from the aggregated safety relevant data collected from a fleet of AVs. Our findings show that the benefits from DyAVSA for the owners/operators of the individual AV instances may be significant: the safety predictions they can make by using the data collected by the entire fleet of AV instances and shared among them, may differ considerably from the predictions the AV instances would be able to make relying on own observations only. Sharing data would lead to a much more rapid reduction of uncertainty in the pfms than would be the case if the AV instances relied on their own observations only. The presented DyAVSA, based on a multivariate Bayesian safety assessment, can be applied to other complex intelligent systems such as robots, UAVs, etc.