Human activity recognition (HAR) systems extract activities from observational data, such as sensor measurements from mobile devices, to provide for instance medical, fitness, or security information. A crucial initial step in these data analysis workflows is segmenting continuous numerical measurements into variable-sized segments that correspond to single activities. Human activity segmentation (HAS) enables downstream classification algorithms to label entire activities. Unfortunately, current time series segmentation (TSS) algorithms exhibit limited performance on multivariate sensor data due to complex temporal dynamics and irrelevant dimensions. This limits their applicability in HAR workflows. In this review, we provide a systematic benchmark of dimensionality reduction, model aggregation, and change point selection applied to the ClaSP TSS algorithm for real-world, multidimensional mobile sensing data. We evaluated the accuracy of the techniques in an experimental study using 250 data sets from the HAS challenge at ECML/PKDD and AALTD 2023. Our findings indicate that extending ClaSP for multivariate data, by aggregating internal representations, yields better results compared to reducing data dimensionality or selecting change points (CPs) from different channels. We report a new state of the art with 73% average accuracy on the challenge benchmark.