Image and Pervasive Access Laboratory

Ensuring the safety of personnel at large industrial sites is a significant challenge. Misunderstandings and increased safety risks are related to communication, as traditional non-targeted alerts like sirens are often ineffective. To address this, we propose a drone-based system that uses voice interaction to deliver timely and targeted safety advisories. This paper presents a multi-stage, user-centered study, beginning with formative interviews with security and safety managers and culminating in an on-site evaluative study with personnel at two biomass power plants. Through our study, we identify key design insights for creating a drone that is perceived as an official, authoritative, and non-threatening entity. We provide design guidance for using sound and movement to signal intent, creating a voice persona that balances authority with clarity, and designing interaction flows to manage confusion and non-compliance. The results demonstrate the system's potential not only to enhance safety compliance, but also to mitigate interpersonal confrontation. First, we provide an empirically-grounded understanding of contextual safety challenges, which informs a set of design principles for intelligible and socially acceptable voice advisories from drones. Second, we present results from an in-situ evaluation of a working prototype that demonstrates its effectiveness in a real-world setting.

Abstract Previous studies on animal models suggested that visual areas involved in motion processing could undergo important cortical reorganizations following retinal damages. This could have major implications for patients suffering from macular degeneration (MD), one leading cause of vision loss. Here, we performed fMRI recordings in a group of maculopathy patients (including individuals suffering from age-related macular degeneration or from Stargardt’s Disease) and a control group to characterize the motion processing cortical network in MD patients and determine whether this network is modified following the onset of the scotoma. We used an experimental protocol based on random-dot kinematograms (RDKs) classically employed to characterize motion-selective areas in the brain. To ensure that the visual information processed by the two groups was equivalent, the visual field in each control participant was masked using an artificial scotoma directly derived from clinical measurements in their paired patient. We found that in MD patients, translational motion elicited significant and robust activations in a restricted cortical network which included the human V5/MT+ complex (hMT+), areas V3A and V6, and a portion of primary visual areas (V1, V2 and V3) connected to peripheral vision. Importantly, the same patterns of responses were also observed in control participants. Moreover, the extent and strength of activation within these motion-selective areas did not differ significantly between the two groups. Altogether, these results suggest that in humans, the motion-selective network does not undergo significant large-scale cortical reorganizations following the onset of MD. Significance statement Motion processing in the visual cortex of patients with macular degeneration has never been characterized. Here, we performed fMRI recordings in 7 maculopathy patients and found robust motion-selective activations in a cortical network which included the human V5/MT+ complex (hMT+), areas V3A and V6, and a portion of primary visual areas connected to peripheral vision. These activations closely align with those reported in participants with normal vision in the literature and do not significantly differ from those measured in a group of age and gender-matched control participants who viewed the motion stimuli with a matched artificial scotoma. Altogether, our results suggest that the motion-selective network does not undergo significant large-scale reorganizations in maculopathy patients following the onset of the scotoma.

Decision Trees (DTs) constitute one of the major highly non-linear AI models, valued, e.g., for their efficiency on tabular data. Learning accurate DTs is, however, complicated, especially for oblique DTs, and does take a significant training time. Further, DTs suffer from overfitting, e.g., they proverbially “do not generalize” in regression tasks. Recently, some works proposed ways to make (oblique) DTs differentiable. This enables highly efficient gradient-descent algorithms to be used to learn DTs. It also enables generalizing capabilities by learning regressors at the leaves simultaneously with the decisions in the tree. Prior approaches to making DTs differentiable rely either on probabilistic approximations at the tree’s internal nodes (soft DTs) or on approximations in gradient computation at the internal node (quantized gradient descent). In this work, we propose DTSemNet, a novel semantically equivalent and invertible encoding for (hard, oblique) DTs as Neural Networks (NNs), that uses standard vanilla gradient descent. Experiments across various classification and regression benchmarks show that oblique DTs learned using DTSemNet are more accurate than oblique DTs of similar size learned using state-of-the-art techniques. Further, DT training time is significantly reduced. We also experimentally demonstrate that DTSemNet can learn DT policies as efficiently as NN policies in the Reinforcement Learning (RL) setup with physical inputs (dimensions ≤32). The code is available at https://github.com/CPS-research-group/dtsemnet.

Abstract We present a super-resolution method for spin systems using a flow-based generative model that is a deep generative model with reversible neural network architecture. Starting from spin configurations on a two-dimensional square lattice, our model generates spin configurations of a larger lattice. As a flow-based generative model precisely estimates the distribution of the generated configurations, it can be combined with Monte Carlo simulation to generate large lattice configurations according to the Boltzmann distribution. Hence, the long-range correlation on a large configuration is reduced into the shorter one through the flow-based generative model. This alleviates the critical slowing down near the critical temperature. We demonstrated an 8 times increased lattice size in the linear dimensions using our super-resolution scheme repeatedly. We numerically show that by performing simulations for 16 × 16 configurations, our model can sample lattice configurations at 128 × 128 on which the thermal average of physical quantities has good agreement with the one evaluated by the traditional Metropolis–Hasting Monte Carlo simulation.

Raised-line graphics are tactile documents made for people with visual impairments (VI). Their exploration relies on a complex two-handed behavior. To better understand the cognitive processes underlying this exploration, we proposed a new method based on “tactile fixations”. A tactile fixation occurs when a finger is stationary within a specific spatial and temporal window. It is known that stationary fingers play an active role when exploring tactile graphics, but they have never been defined or studied before. In this study, we first defined the concept of tactile fixation, then we conducted a behavioral study with ten participants with VI in order to assess the role of tactile fixations under different conditions. The results show that tactile fixations vary according to different factors such as the graphic type as well as the involved hand and the aim of the exploration.

Visually impaired people encounter many challenges in their everyday life, especially when it comes to navigating and representing space. The issue of shopping is addressed mostly on the level of navigation and product detection, but conveying clues about the object position to the user is rarely implemented. This work presents a prototype of vibrotactile wristband using spatiotemporal patterns to help visually impaired users reach an object in the 2D plane in front of them. A pilot study on twelve blindfolded sighted subjects showed that discretizing space in a seven by seven targets matrix and conveying clues with a discrete pattern on the vertical axis and a continuous pattern on the horizontal axis is an intuitive and effective design.

International audience

Internet Of Things (IoT) is increasingly used throughout the world in different fields. But it does not have a standardized definition [28]. Several definition can be proposed. IoT corresponds to “objects with virtual identity and personality, working in a smart environment and using smart interfaces to connect and communicate in some various context” [3]. IoT is a sum of entities that are used to exchange information in different contexts. This is a network of connected objects communicating between them to extend their functionalities [17]. The IoT is larger than just a system: it is a system of system. Each one can be divided in sub-system and assimilated to a specific technology [5]. The IoT is “a dynamic infrastructure of a global network. This global network has auto-configuration capacities based on standards and communication protocols interoperable. In this network, physical and virtual objects have identities, physical attributes, virtual personalities and smart interfaces and they are integrated to the network in a transparent way” [31].

This paper presents an experiment that shows the impact of the transitions of smartphone positions (from one position to another on the human body) on the recognition of the human activity, in an uncontrolled environment. The studied activities are walking, laying, standing and sitting. In this environment, users are free to switch smartphone position, for instance from hand to pocket. Our methodology includes the combination of different sensors and a classification with a Deep Neural Network algorithm. Two datasets were considered, the only difference between the two datasets is that the transition are removed in the second dataset. Results show an improvement of accuracy of the classification of activities 93.87% for the control dataset compare to 95.07% for the dataset with removed transitions. We also show that with no transitions the convergence is quicker and more robust. Consequently, our method including removing information corresponding to transitions should save memory space and computing time while offering a high classification precision.