International audience

Understanding and classifying mobility modes, such as walking, cycling, driving, or public transport, is essential for sustainable urban planning and mobility behavior analysis. Traditional approaches rely on handcrafted features and machine learning models trained on GPS trajectory data. However, these methods require extensive data preparation and model training. In this work, we explore the potential of large language models (LLMs) as zero-shot predictors for transportation mode classification, eliminating the need for training data altogether. We propose a pipeline that transforms enriched trajectory segments into textual prompts, enabling LLMs to perform classification without task-specific pretraining. We benchmark the performance of a locally distilled 32B parameter LLM (DeepSeek Gwen) against standard machine learning baselines on the Geolife dataset. Preliminary results demonstrate that LLMs effectively capture semantic and contextual cues from trajectory-derived features, highlighting their promise for rapid, data-efficient transportation mode classification. Our work provides novel insights into leveraging LLMs in the mobility domain and identifies future opportunities for their integration.

This paper focuses on GNSS trajectory aggregation, building upon the work of [4], a modular and iterative aggregation algorithm. The last allows to compute aggregated trajectories with high geometric accuracy for a set of trajectories having the same origin/destinations and following the same path. To analyze the performances of the algorithm, we use a metrological perspective, i.e. its ability to reconstruct the accurate aggregated trajectory, from a minimum number of input GNSS trajectory samples to determine an optimal parameterization of the algorithm. We validate our modular framework on two types of data: synthetic and experimental GNSS trajectories collected with multiple sensors under varying canopy conditions.

Land cover mapping is crucial for natural resource assessment, urban planning, and sustainable development. Land cover nomenclature often includes two or three hierarchic levels with tree-like hierarchical structures. This study aims to explore these hierarchical relationships and the potential of hierarchical semantic segmentation for land cover mapping. We propose a hierarchical semantic segmentation architecture by taking advantage of dual U-shape network, named as HierU-Net. The coarse-level result is ingested to the fine-level segmentation functioned as soft constraints. The propagation of error will not be certain. Moreover, we employ a multi-task loss function weighted by homoscedastic uncertainty to optimize the training. To evaluate the performance of the proposed method, we create a hierarchical semantic segmentation dataset (HierToulouse), which contains 11,528 samples, including images and land cover labels at two hierarchical levels. The experiments demonstrate that the proposed approach is capable of achieving accurate land cover segmentation at both coarse and fine levels, with segmentation results surpassing those obtained using the flat method.

In this paper, we examine the properties of the radial distance which has been used as a tool to compare the shape of simple surfacic objects. We give a rigorous definition of the radial distance and derive its theoretical properties, and in particular under which conditions it satisfies the distance properties. We show how the computation of the radial distance can be implemented in practice and made faster by the use of an analytical formula and a Fast Fourier Transform. Finally, we conduct experiments to measure how the radial distance is impacted by perturbation and generalization and we give abacuses and thresholds to deduce when buildings are likely to be homologous or non-homologous given their radial distance.