M. Badii\u00a0JOUABER<\/span><\/b>, Professor, T\u00e9l\u00e9com SudParis, FRANCE – Directeur de theseMme Salima\u00a0BENBERNOU<\/span><\/b>, Professor, Universit\u00e9 Paris Cit\u00e9, FRANCE – ExaminateurM. Ken\u00a0CHEN<\/span><\/b>, Professor, Universit\u00e9 Paris 13, FRANCE – ExaminateurM. Nadjib\u00a0AIT SAADI<\/span><\/b>, Professor, UVSQ Paris-Saclay \/ Paris-Saclay university , FRANCE – RapporteurM. Pascal\u00a0LORENZ<\/span><\/b>, Professor, Universit\u00e9 de Haute-Alsace, FRANCE – RapporteurM. Hassine\u00a0MOUNGLA<\/span><\/b>, Professor, Universit\u00e9 Paris Cit\u00e9, FRANCE – Co-encadrant de these<\/p> L’arriv\u00e9e The L’Ecole doctorale : Ecole Doctorale de l’Institut Polytechnique de Pariset le Laboratoire de recherche SAMOVAR – Services r\u00e9partis, Architectures, Mod\u00e9lisation, Validation, Administration des R\u00e9seaux pr\u00e9sentent l\u2019AVIS DE SOUTENANCE de Monsieur Mehdi Salim BENHELAL Autoris\u00e9 \u00e0 pr\u00e9senter ses travaux en vue de l\u2019obtention du Doctorat de l’Institut Polytechnique de Paris, pr\u00e9par\u00e9 \u00e0 T\u00e9l\u00e9com SudParis en : […]<\/p>\n","protected":false},"author":4,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ocean_post_layout":"","ocean_both_sidebars_style":"","ocean_both_sidebars_content_width":0,"ocean_both_sidebars_sidebars_width":0,"ocean_sidebar":"","ocean_second_sidebar":"","ocean_disable_margins":"enable","ocean_add_body_class":"","ocean_shortcode_before_top_bar":"","ocean_shortcode_after_top_bar":"","ocean_shortcode_before_header":"","ocean_shortcode_after_header":"","ocean_has_shortcode":"","ocean_shortcode_after_title":"","ocean_shortcode_before_footer_widgets":"","ocean_shortcode_after_footer_widgets":"","ocean_shortcode_before_footer_bottom":"","ocean_shortcode_after_footer_bottom":"","ocean_display_top_bar":"default","ocean_display_header":"default","ocean_header_style":"","ocean_center_header_left_menu":"","ocean_custom_header_template":"","ocean_custom_logo":0,"ocean_custom_retina_logo":0,"ocean_custom_logo_max_width":0,"ocean_custom_logo_tablet_max_width":0,"ocean_custom_logo_mobile_max_width":0,"ocean_custom_logo_max_height":0,"ocean_custom_logo_tablet_max_height":0,"ocean_custom_logo_mobile_max_height":0,"ocean_header_custom_menu":"","ocean_menu_typo_font_family":"","ocean_menu_typo_font_subset":"","ocean_menu_typo_font_size":0,"ocean_menu_typo_font_size_tablet":0,"ocean_menu_typo_font_size_mobile":0,"ocean_menu_typo_font_size_unit":"px","ocean_menu_typo_font_weight":"","ocean_menu_typo_font_weight_tablet":"","ocean_menu_typo_font_weight_mobile":"","ocean_menu_typo_transform":"","ocean_menu_typo_transform_tablet":"","ocean_menu_typo_transform_mobile":"","ocean_menu_typo_line_height":0,"ocean_menu_typo_line_height_tablet":0,"ocean_menu_typo_line_height_mobile":0,"ocean_menu_typo_line_height_unit":"","ocean_menu_typo_spacing":0,"ocean_menu_typo_spacing_tablet":0,"ocean_menu_typo_spacing_mobile":0,"ocean_menu_typo_spacing_unit":"","ocean_menu_link_color":"","ocean_menu_link_color_hover":"","ocean_menu_link_color_active":"","ocean_menu_link_background":"","ocean_menu_link_hover_background":"","ocean_menu_link_active_background":"","ocean_menu_social_links_bg":"","ocean_menu_social_hover_links_bg":"","ocean_menu_social_links_color":"","ocean_menu_social_hover_links_color":"","ocean_disable_title":"default","ocean_disable_heading":"default","ocean_post_title":"","ocean_post_subheading":"","ocean_post_title_style":"","ocean_post_title_background_color":"","ocean_post_title_background":0,"ocean_post_title_bg_image_position":"","ocean_post_title_bg_image_attachment":"","ocean_post_title_bg_image_repeat":"","ocean_post_title_bg_image_size":"","ocean_post_title_height":0,"ocean_post_title_bg_overlay":0.5,"ocean_post_title_bg_overlay_color":"","ocean_disable_breadcrumbs":"default","ocean_breadcrumbs_color":"","ocean_breadcrumbs_separator_color":"","ocean_breadcrumbs_links_color":"","ocean_breadcrumbs_links_hover_color":"","ocean_display_footer_widgets":"default","ocean_display_footer_bottom":"default","ocean_custom_footer_template":"","ocean_post_oembed":"","ocean_post_self_hosted_media":"","ocean_post_video_embed":"","ocean_link_format":"","ocean_link_format_target":"self","ocean_quote_format":"","ocean_quote_format_link":"post","ocean_gallery_link_images":"on","ocean_gallery_id":[],"footnotes":""},"categories":[286,402],"tags":[],"class_list":["post-6955","post","type-post","status-publish","format-standard","hentry","category-fractualites-ennews-fr","category-seminaires-ness-2013-fr","entry"],"_links":{"self":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/6955","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/comments?post=6955"}],"version-history":[{"count":1,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/6955\/revisions"}],"predecessor-version":[{"id":6956,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/6955\/revisions\/6956"}],"wp:attachment":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/media?parent=6955"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/categories?post=6955"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/tags?post=6955"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}\u00ab
Architectures intelligentes de nouvelle g\u00e9n\u00e9ration pour des v\u00e9hicules
autonomes et connect\u00e9s fiables : approches utilisant l\u2019assistance edge
et l\u2019apprentissage f\u00e9d\u00e9r\u00e9 \u00bb<\/h2>pr\u00e9sent\u00e9 par Monsieur Mehdi Salim BENHELAL<\/h2><\/center>R\u00e9sum\u00e9 :<\/b>
des v\u00e9hicules connect\u00e9s et autonomes (CAVs) est responsable de la
r\u00e9volution du transport des personnes et des biens. N\u00e9anmoins, il est
n\u00e9cessaire d’avoir un syst\u00e8me de pr\u00e9diction de trajectoire pr\u00e9cis et
robuste pour atteindre une autonomie qui est fiable, celui-ci constitue
une passerelle entre la perception et la planification d’actions.
Cependant, un tel contexte soul\u00e8ve de nombreux d\u00e9fis li\u00e9s aux
contraintes de communication, \u00e0 la pr\u00e9servation de la vie priv\u00e9e, ainsi
qu’aux exigences d’exactitude et de robustesse. Nous abordons ces
probl\u00e9matiques \u00e0 travers trois cadres compl\u00e9mentaires d\u00e9di\u00e9s \u00e0 la
pr\u00e9diction de trajectoires dans les CAVs. Tout d’abord, nous pr\u00e9sentons
une nouvelle architecture de clustering assist\u00e9e par l’edge qui utilise
des mod\u00e8les d’apprentissage profond et des technologies de edge pour
am\u00e9liorer les pr\u00e9visions. Les mod\u00e8les de base de tous les CAVs prennent
en entr\u00e9e les positions historiques de leur v\u00e9hicule cible pour g\u00e9n\u00e9rer
des pr\u00e9dictions, qui sont envoy\u00e9es \u00e0 un serveur edge. Les trajectoires
pr\u00e9vues sont ensuite regroup\u00e9es \u00e0 l’aide de l’algorithme DBSCAN en
partitions de trajectoires, o\u00f9 les trajectoires du plus grand cluster
sont moyenn\u00e9es et le r\u00e9sultat renvoy\u00e9 \u00e0 tous les CAVs. Nous constatons
que notre architecture obtient d’excellentes performances sur le jeu<\/span><\/span>
de donn\u00e9es nuScenes et r\u00e9siste aux probl\u00e8mes de capacit\u00e9 pr\u00e9dictive
compromise d’un seul agent. La faisabilit\u00e9 du syst\u00e8me est \u00e9galement
analys\u00e9e avec les capacit\u00e9s 5G\/6G existantes. Deuxi\u00e8mement, nous
proposons une approche d’apprentissage f\u00e9d\u00e9r\u00e9 (FL) pour la pr\u00e9diction de
trajectoire. Cette approche int\u00e8gre un r\u00e9seau neuronal siamois (SNN)
afin d’identifier les similarit\u00e9s contextuelles entre les environnements
clients. Les clients pr\u00e9sentant des contextes statiques similaires sont
regroup\u00e9s avant l’apprentissage f\u00e9d\u00e9r\u00e9, ce qui am\u00e9liore la
collaboration et les performances de l’architecture. Des exp\u00e9riences sur
un jeu<\/span><\/span>
de donn\u00e9es personnalis\u00e9 utilisant des sc\u00e8nes de highway drone dataset
(highD) et intersection drone dataset (inD) \u00e9valu\u00e9es avec l’Average
Displacement Error (ADE) et la Final Displacement Error (FDE),
d\u00e9montrent l’efficacit\u00e9 de cette m\u00e9thode pour la pr\u00e9diction de
trajectoires en conditions r\u00e9elles tout en pr\u00e9servant la
confidentialit\u00e9. Troisi\u00e8mement, nous mettons en avant une nouvelle
architecture de FL d\u00e9centralis\u00e9e sur plusieurs niveaux pour la
pr\u00e9diction de trajectoire. Cette conception r\u00e9sout les probl\u00e8mes de
point de d\u00e9faillance unique (SPOF), de goulots d’\u00e9tranglement, de
communication et de d\u00e9pendance \u00e0 un serveur central impactant
n\u00e9gativement l’apprentissage f\u00e9d\u00e9r\u00e9 centralis\u00e9 (CFL). Nous cr\u00e9ons un
framework d\u00e9centralis\u00e9 de deux niveaux et compos\u00e9 de plusieurs n\u0153uds,
chaque n\u0153ud repr\u00e9sentant un client unique. Les clients utilisent
principalement la communication parent-enfant, ce qui diminue le co\u00fbt de
transmission. Nous d\u00e9montrons que notre m\u00e9thode atteint une pr\u00e9cision
de pr\u00e9diction comparable \u00e0 celle du CFL tout en r\u00e9duisant les co\u00fbts de
communication sur les jeux de donn\u00e9es highD et inD. Nos architectures
repr\u00e9sentent un progr\u00e8s consid\u00e9rable en mati\u00e8re de pr\u00e9diction de
trajectoire au sein des CAVs. Dans ce travail de th\u00e8se, plusieurs d\u00e9fis
sont relev\u00e9s en termes de pr\u00e9cision, de confidentialit\u00e9, de robustesse
et de scalabilit\u00e9 gr\u00e2ce aux m\u00e9thodes assist\u00e9es par l’edge, de FL
contextuel et de d\u00e9centralisation hi\u00e9rarchique. Ce travail constitue une
base pour d\u00e9velopper des solutions de pr\u00e9diction de trajectoire plus
robustes et mieux adapt\u00e9es aux contextes r\u00e9els de conduite autonome.<\/p>Abstract :<\/b>
advent of Connected and Autonomous Vehicles (CAVs) plays a massive role
in people and goods transportation. To attain reliable autonomy, a
precise and robust trajectory prediction system serving as a connector
between perception and action planning is required. However, many
problems related to the constraints of communication, privacy
preservation, accuracy, and robustness arise in such a setting. We
tackle these issues through three complementary frameworks for
trajectory prediction in CAVs ecosystem. First, we present a new
edge-assisted clustering architecture that utilizes deep learning models
and edge technology to enhance forecasting. The base models of all CAVs
take as input the historical locations of their target vehicle to
generate predictions, which are sent to an edge server. The forecasted
trajectories are then clustered using DBSCAN algorithm into trajectory
partitions, where the largest cluster is averaged and sent back to all
CAVs. We find that our architecture obtains strong performance on the
nuScenes dataset and is resilient to a single agent\u2019s compromised
predictive capability. The system feasibility is also analyzed with
existing 5G\/6G capabilities. Second, we propose a Federated Learning
(FL) approach for trajectory prediction that incorporates a Siamese
Neural Network (SNN) to identify contextual similarities between client
environments. Clients with similar static contexts are clustered
together before the federated training, leading to more efficient
collaboration and improved architecture performance. Experiments on a
custom dataset using scenes from the highway drone dataset (highD) and
intersection drone dataset (inD), evaluated with Average Displacement
Error (ADE) and Final Displacement Error (FDE), demonstrate that this
method is effective for real-world trajectory prediction while
preserving privacy. Third, we introduce a new multi-level decentralized
Federated Learning architecture for trajectory prediction. This design
solves the problems of single point of failure (SPOF), communication
bottlenecks, and reliance on a central server that Centralized Federated
Learning (CFL) suffers from. We create our two-level decentralized
framework composed of multiple nodes where each node represents a single
client. Clients mainly have parent-child communication, which limits
the transmission overhead. We show that our method achieves comparable
prediction accuracy to CFL while reducing communication costs on the
highD and inD datasets. Our frameworks represent a significant
advancement in trajectory prediction for the CAV ecosystem. This thesis
tackles important challenges by providing solutions that target
accuracy, privacy, robustness, and scalability. To this end, we design
edge-assisted consensus, context-aware FL, and hierarchical
decentralization methods. This work provides a basis for trajectory
prediction methods that are not only more reliable but also more closely
aligned with real-world applications in autonomous driving.<\/p>