International audience

International audience

Spiking Neural Networks (SNNs) are increasingly explored for their energy efficiency and biological plausibility, offering a compelling alternative to conventional artificial neural networks in neuromorphic applications. However, even Spik-former a fully spiking adaptation of the Transformer model, can exhibit significant computational redundancy due to excessive spike activity, resulting in non-negligible energy consumption. In this paper, we introduce Masked Spikformer, a unified framework for regulating temporal spike activity in spiking Transformers through three complementary masking strategies: Random Spike Masking (RSM), Gaussian-Based Spike Masking (GSM), and Gaussian-Based Spike Weighting (GSW). These approaches encompass both binary masking and continuous, learnable temporal weighting. Our method is integrated into fully spiking architectures and applied consistently during both training and inference. Experimental results on neuromorphic benchmarks (CIFARIO-DVS and DVS128 Gesture) show that the proposed masking strategies significantly reduce energy consumption while maintaining high classification performance. Notably, the best accuracy on DVS128 Gesture is achieved by RSM with 70% masking, while GSW(i), the inverse Gaussian weighting variant, attains the highest accuracy on CIFARIO-DVS. RSM also provides the lowest energy consumption across both datasets, highlighting the effectiveness of temporal sparsity for energy-efficient spiking Transformers.

This paper provides a comprehensive review of the integration of Spiking Neural Networks (SNNs) and Transformers, combining the energy efficiency of SNNs with the high performance of Transformer architectures. By leveraging the event-driven nature of SNNs and the powerful self-attention mechanism of Transformers, this fusion aims to address the challenges of high energy consumption in deep learning while improving task accuracy, especially for complex datasets. We introduce the core concepts of SNNs and Transformers, reviewing state-of-the-art methods for their combination, including hybrid architectures. The performance of each architecture is presented thanks to both static and neuromorphic datasets, highlighting their advantages and limitations. This review also discusses the challenges of integrating self-attention into spiking architectures and outlines future research directions to further enhance model performance and energy efficiency.

Recently the use of Unmanned Aerial Vehicles (UAVs), commonly referred to as drones, has exploded in several domains, leading to potential security issues. As such, estimating the exact position of those eventual malicious drones has become of crucial interest. However, computing an accurate and precise geolocation of these drones, especially in outdoor environments, remains challenging. This paper focuses on outdoor 3-dimensional (3D) drone geolocation techniques based on Radio Frequency (RF) signals. We first present a RF-based 3D drone geolocation dataset, and then apply and compare various geolocation techniques, ranging from geometrical-based to machine learning-based methods. We further propose a new hybrid method blending the two above categories of geolocation techniques, that achieves an average 3D error of the order of 11.7 meters within a search volume of about 520×560×115 m3, significantly below the one achieved with geometrical-based techniques, and with a reduced computational complexity compared to the regular machine-learning based techniques.

Gesture recognition is a tool to enable novel interactions with different techniques and applications, like Mixed Reality and Virtual Reality environments. With all the recent advancements in gesture recognition from skeletal data, it is still unclear how well state-of-the-art techniques perform in a scenario using precise motions with two hands. This paper presents the results of the SHREC 2024 contest organized to evaluate methods for their recognition of highly similar hand motions using the skeletal spatial coordinate data of both hands. The task is the recognition of 7 motion classes given their spatial coordinates in a frame-by-frame motion. The skeletal data has been captured using a Vicon system and pre-processed into a coordinate system using Blender and Vicon Shogun Post. We created a small, novel dataset with a high variety of durations in frames. This paper shows the results of the contest, showing the techniques created by the 5 research groups on this challenging task and comparing them to our baseline method.

Abstract Falls are a significant health risk for older adults, and fall-related injuries are a leading cause of morbidity and mortality in this population. Elderlies are particularly vulnerable to falls due to age-related declines in mobility, balance, and muscle strength, as well as chronic medical conditions with use of certain medications. These injuries can range from minor bruises and scrapes to more severe like fractures, head trauma, or internal bleeding. To prevent falls in older adults, some solutions propose to ensure a safe living environment, others to maintain physical activity, and others to manage chronic medical conditions. This article presents the implementation and test of a system preventing hip fractures resulting from falls using a fall detection and prediction system designed to protect and alert individuals during falls.

In the field of train transportation, personal injuries due to train automatic doors are still a common occurrence. This paper aims at implementing a computer vision solution as part of a safety detection system to identify automatic doors-related hazardous events to reduce their occurrence and their severity. Deep anomaly detection algorithms are often applied on CCTV video feeds to identify such hazardous events. However, the anomalous events identified by those algorithms are often simpler than most common occurrences in transport environments, hindering their widespread usage. Since such events are of quite a diverse nature and no dataset featuring them exist, we create a specilically-tailored dataset composed of real-case scenarios of hazardous events near train doors. We then study an anomaly detection algorithm from the literature on this dataset and propose a set of modifications to better adapt it to our railway context and to subsequently ease its application to a wider range of use-cases.

In this work, we evaluated the performances of an adaptive and online clustering algorithm (Self-Adaptive Kernel Machine - SAKM) adjusted for the automatic and online recognition of attacks on wi-fi communication (802.11n protocol). The results presented here are part of a wider project dealing with wi-fi system monitoring. The radio waves are easy to listen. Due to the quick evolution in the available attacks, the use of learning algorithm cannot cover all configurations. Online clustering constructs evolving models without knowledge of the different cases to discriminate and is therefore well suited to this type of problematic. Based on SVM and kernel methods, the SAKM algorithm uses a fast adaptive learning procedure to take into account variations over time.

This paper introduces a novel approach for early recognition of human actions using 3D skeleton joints extracted from 3D depth data. We propose a novel, frame-by-frame and real-time descriptor called Body-part Directional Velocity (BDV) calculated by considering the algebraic velocity produced by different body-parts. A real-time Hidden Markov Models algorithm with Gaussian Mixture Models state-output distributions is used to carry out the classification. We show that our method outperforms various state-of-the-art skeleton-based human action recognition approaches on MSRAction3D and Florence3D datasets. We also proved the suitability of our approach for early human action recognition by deducing the decision from a partial analysis of the sequence.

An online recognition system must analyze the changes in the sensing data and at any significant detection; it has to decide if there is a change in the activity performed by the person. Such a system can use the previous sensor readings for decision-making (decide which activity is performed), without the need to wait for future ones. This paper proposes an approach of human activity recognition on online sensor data. We present four methods used to extract features from the sequence of sensor events. Our experimental results on public smart home data show an improvement of effectiveness in classification accuracy.

By 2050, about one third of the French population will be over 65. Our laboratory's current research focuses on the monitoring of elderly people at home, to detect a loss of autonomy as early as possible. Our aim is to quantify criteria such as the international activities of daily living (ADL) or the French Autonomie Gerontologie Groupes Iso-Ressources (AGGIR) scales, by automatically classifying the different ADL performed by the subject during the day. A Health Smart Home is used for this. Our Health Smart Home includes, in a real flat, infrared presence sensors (location), door contacts (to control the use of some facilities), temperature and hygrometry sensor in the bathroom, and microphones (sound classification and speech recognition). A wearable kinematic sensor also informs postural transitions (using pattern recognition) and walk periods (frequency analysis). This data collected from the various sensors are then used to classify each temporal frame into one of the ADL that was previously acquired (seven activities: hygiene, toilet use, eating, resting, sleeping, communication, and dressing/undressing). This is done using support vector machines. We performed a 1-h experimentation with 13 young and healthy subjects to determine the models of the different activities, and then we tested the classification algorithm (cross validation) with real data.

A biofeedback system whose underlying principle consists in supplying the user with supplementary sensory information related to foot sole pressure distribution through a tongue-placed tactile output device was recently developed for improving balance. The purpose of this study was to unravel the underlying control mechanisms involved in the regulation of the centre of foot pressure (CoP) displacements during upright quiet standing with this biofeedback system. Ten young healthy adults were asked to stand as immobile as possible with their eyes closed in two conditions of No-biofeedback and Biofeedback. CoP displacements, recorded using a force platform, were processed through a space-time domain analysis and modeled as fractional Brownian motions according to the procedure of stabilogram-diffusion analysis (SDA). The space-time domain analysis showed decreased CoP displacements in the Biofeedback relative to No-biofeedback condition. Complementary, the SDA showed decreased spatiotemporal threshold at which corrective mechanisms are called into play and an increased degree of control of the CoP displacements in the Biofeedback relative to No-biofeedback condition. The present findings evidence that the effectiveness of the biofeedback in decreasing the CoP displacements during upright quiet standing stems from an increased contribution and efficiency of anti-persistent mechanisms (feedback control) involved in the regulation of the CoP displacements.

4p.