https:\/\/zoom.us\/j\/2445226856<\/a><\/p>\n\n\n\nMembres du jury :<\/strong><\/p>\n\n\n\nM. Noel CRESPI<\/strong>, Professor, Institut Polytechnique de Paris T\u00e9l\u00e9com SudParis, FRANCE – Directeur de these
Mme Rajapksha PRABODA<\/strong>, Ma\u00eetre de recherche, Institut Polytechnique de Paris T\u00e9l\u00e9com SudParis, FRANCE – Co-encadrant de these
M. Shiping WANG<\/strong>, Professor, Fuzhou university, CHINE – Rapporteur
M. Ioan Marius BILASCO<\/strong>, Ma\u00eetre de conf\u00e9rences, Universit\u00e9 de Lille, FRANCE – Rapporteur
M. Gu BIN<\/strong>, Ma\u00eetresse de conf\u00e9rences, Mohamed bin Zayed University of AI, ARABIE SAOUDITE – Examinateur
Mme Patricia DESGREYS<\/strong>, Professor, Institut Polytechnique de Paris Telecom Paris, FRANCE – Examinateur<\/p>\n\n\n\n\u00ab Am\u00e9lioration de la d\u00e9tection d’anomalies vid\u00e9o bas\u00e9e sur des technique avanc\u00e9es d’Apprentissage Profond \u00bb<\/h2>\n\n\n\npr\u00e9sent\u00e9 par Monsieur Wenhao SHAO<\/h2>\n\n\n\n
R\u00e9sum\u00e9 :<\/strong><\/p>\n\n\n\nLa s\u00e9curit\u00e9 est une pr\u00e9occupation majeure dans diff\u00e9rents domaines, et le d\u00e9ploiement de syst\u00e8mes de surveillance en temps r\u00e9el permet de relever ce d\u00e9fi. En utilisant des techniques d’apprentissage profond, il permet de reconna\u00eetre efficacement les \u00e9v\u00e9nements anormaux. Cependant, m\u00eame avec les avanc\u00e9es actuelles des m\u00e9thodes de d\u00e9tection des anomalies, distinguer les \u00e9v\u00e9nements anormaux des \u00e9v\u00e9nements normaux dans les sc\u00e9narios du monde r\u00e9el reste un d\u00e9fi en raison d\u2019\u00e9v\u00e9nements anormaux rares, visuellement diversifi\u00e9s et non reconnaissables de fa\u00e7on pr\u00e9visible. Cela est particuli\u00e8rement vrai lorsque l’on s’appuie sur des m\u00e9thodes supervis\u00e9es, o\u00f9 le manque de donn\u00e9es d’anomalies labelis\u00e9es pose un probl\u00e8me important pour distinguer les vid\u00e9os normales des vid\u00e9os anormales. Par cons\u00e9quent, les approches de d\u00e9tection d’anomalies les plus r\u00e9centes utilisent des ensembles de donn\u00e9es existants pour concevoir ou apprendre un mod\u00e8le qui capture les mod\u00e8les normaux, ce qui permet ensuite d’identifier les mod\u00e8les anormaux inconnus. Au cours de la phase de conception du mod\u00e8le, il est essentiel de labelliser les vid\u00e9os avec des attributs tels qu’une apparence anormale, un comportement ou des cat\u00e9gories cibles qui s’\u00e9cartent de mani\u00e8re significative des donn\u00e9es normales, en les marquant comme des anomalies. Outre le manque de donn\u00e9es labellis\u00e9es, trois autres d\u00e9fis principaux ont \u00e9t\u00e9 identifi\u00e9s dans la litt\u00e9rature : 1) la repr\u00e9sentation insuffisante des caract\u00e9ristiques temporelles, 2) le manque de pr\u00e9cision dans le positionnement des \u00e9v\u00e9nements anormaux et 3) l’absence d’informations sur le comportement. Nous avons explor\u00e9 les applications des nouvelles technologies de traitement vid\u00e9o, notamment la reconnaissance des actions, la d\u00e9tection des cibles, l’extraction des caract\u00e9ristiques du flux optique, l’apprentissage de la repr\u00e9sentation et l’apprentissage contrastif, afin de les utiliser dans les mod\u00e8les de d\u00e9tection des anomalies vid\u00e9o. Les mod\u00e8les que nous proposons sont analys\u00e9s de mani\u00e8re comparative avec les mod\u00e8les de r\u00e9f\u00e9rence. Cette analyse comparative est r\u00e9alis\u00e9e \u00e0 l’aide de jeux de donn\u00e9es publics courants, notamment UCSD(Ped2), Avenue, UCF-Crime et Shanghaitech. La premi\u00e8re contribution rel\u00e8ve le premier point d\u00e9crit ci-dessus en introduisant un r\u00e9seau convolutionnel temporel (TCN) am\u00e9lior\u00e9. Ce nouveau mod\u00e8le de r\u00e9seau convolutionnel temporel apprend les caract\u00e9ristiques dynamiques de la vid\u00e9o et les optimise afin d’att\u00e9nuer les erreurs dues aux poids initiaux appris de mani\u00e8re contrastive. Cette m\u00e9thode am\u00e9liore la capacit\u00e9 globale des mod\u00e8les faiblement supervis\u00e9s en r\u00e9duisant la perte caus\u00e9e par les param\u00e8tres initiaux dans l’apprentissage contrastif. N\u00e9anmoins, l’apprentissage faiblement supervis\u00e9 ne fait que r\u00e9duire la d\u00e9pendance \u00e0 l’\u00e9gard des donn\u00e9es labellis\u00e9es, mais ne l’\u00e9limine pas compl\u00e8tement. C’est pourquoi nos deux contributions suivantes s’appuient sur l’apprentissage non supervis\u00e9 pour relever les deux autres d\u00e9fis mentionn\u00e9s ci-dessus. La deuxi\u00e8me contribution combine le m\u00e9canisme d’auto-attention pour donner la priorit\u00e9 aux poids des zones pr\u00e9sentant des fluctuations dynamiques \u00e9videntes dans les images. Lors des tests, les zones anormales sont localis\u00e9es en comparant les fonctions de d\u00e9tection et de perte d’objets. La troisi\u00e8me contribution explore l’int\u00e9gration de mod\u00e8les de r\u00e9seaux d’apprentissage collaboratifs, qui assurent la coh\u00e9rence entre les informations sur le flux optique et les informations sur l’apparence. Cette int\u00e9gration vise \u00e0 am\u00e9liorer les capacit\u00e9s de capture spatio-temporelle des mod\u00e8les non supervis\u00e9s. Les performances et les capacit\u00e9s globales du mod\u00e8le non supervis\u00e9 sont consid\u00e9rablement am\u00e9lior\u00e9es par rapport aux autres mod\u00e8les de base.
<\/p>\n\n\n\n
Abstract :<\/strong><\/p>\n\n\n\nSecurity in public spaces is a primary concern across different domains and the deployment of real-time monitoring systems addresses this challenge. Video surveillance systems employing deep learning techniques allows for the effective recognition of anomaly events. However, even with the current advances in anomaly detection methods, distinguishing abnormal events from normal events in real-world scenarios remains a challenge because they often involve rare, visually diverse, and unrecognizable abnormal events. This is particularly true when relying on supervised methods, where the lack of sufficient labeled anomaly data poses a significant challenge for distinguishing between normal and abnormal videos. As a result, state-of-the-art anomaly detection approaches utilize existing datasets to design or learn a model that captures normal patterns, which is then helpful in identifying unknown abnormal patterns. During the model design stage, it is crucial to label videos with attributes such as abnormal appearance, behavior, or target categories that deviate significantly from normal data, marking them as anomalies. In addition to the lack of labeled data, we identified three challenges from the literature: 1) insufficient representation of temporal feature, 2) lack of precise positioning of abnormal events and 3) lack the consistency research of temporal feature and appearance feature. The objective of my thesis is to propose and investigate advanced video anomaly detection methods by addressing the aforementioned challenges using novel concepts and utilizing weak supervision and unsupervised models rather than relying on supervised models. We actively explored the applications of new video processing technologies, including action recognition, target detection, optical flow feature extraction, representation learning, and contrastive learning in order to utilize them in video anomaly detection models. Our proposed models comparatively analysed with baseline models. This comparative analysis are conducted using prevalent public datasets, including UCSD(Ped2), Avenue, UCF-Crime, and Shanghaitech. The first contribution addresses the first challenge outlined above by introducing an enhanced Temporal Convolutional Network (TCN). This novel TCN model learns dynamic video features and optimizes features to mitigate errors due to contrastive learned initial weights. This method enhances the overall capability of weakly supervised models by reducing the loss caused by initial parameters in contrastive learning. Nevertheless, weakly supervised learning only reduces the reliance on labeled data but does not eliminate the dependence on such data. Hence, our subsequent two contributions rely on unsupervised learning to addressing the other two challenges mentioned above. The second contribution combines the self-attention mechanism to prioritize the weights of areas with obvious dynamic fluctuations in frames. And, during the testing, abnormal areas are located through comparison of object detection and loss functions. The combination of self-attention mechanism and object detection significantly improves the detection accuracy and expands the functionality. The third contribution explores the integration of collaborative teaching network models, which bridges consistency between optical flow information and appearance information. This integration aims to enhance the spatio-temporal capture capabilities of unsupervised models. The overall performance and capabilities of the unsupervised model are significantly enhanced compared to the other baseline models.<\/p>\n","protected":false},"excerpt":{"rendered":"
L’Ecole doctorale : Ecole Doctorale de l’Institut Polytechnique de Paris et le Laboratoire de recherche SAMOVAR – Services r\u00e9partis, Architectures, Mod\u00e9lisation, Validation, Administration des R\u00e9seaux pr\u00e9sentent l\u2019AVIS DE SOUTENANCE de Monsieur Wenhao SHAO 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":"0","ocean_second_sidebar":"0","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":"0","ocean_custom_header_template":"0","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":"0","ocean_menu_typo_font_family":"0","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":"0","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":"off","ocean_gallery_id":[],"footnotes":""},"categories":[286,402],"tags":[],"class_list":["post-6391","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\/6391","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=6391"}],"version-history":[{"count":1,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/6391\/revisions"}],"predecessor-version":[{"id":6392,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/6391\/revisions\/6392"}],"wp:attachment":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/media?parent=6391"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/categories?post=6391"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/tags?post=6391"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}