La production des circulations ferroviaires a la sncf repose actuellement sur un processus essentiellement sequentiel dans lequel la conception des grilles horaires de circulation (reservation de l'infrastructure pour la circulation des trains de l'offre de transport de la sncf) conditionne largement la conception des planifications des engins ferroviaires (les roulements engins), puis celle des agents de conduite (adc) (les grilles de service des adc). cette strategie de planification sequentielle des ressources ferroviaires a ete massivement adoptee pour des raisons pratiques et scientifiques (historique, savoir-faire, complexite du systeme ferroviaire, etc.). toutefois, cette strategie de planification sequentielle genere des solutions qui peuvent etre de cout eleve et moins robustes aux aleas, car les decisions prises a une etape donnee peuvent reduire considerablement l'ensemble des solutions realisables aux etapes suivantes. face a ce constat et a la forte interaction entre ces trois ressources heterogenes et tres couteuses, la sncf a souhaite investiguer la praticabilite et les apports d'une demarche d'optimisation du plan de transport par planification integree de ces ressources critiques. dans cette optique, les travaux de these ont porte sur l'etude de faisabilite, le prototypage et la validation d'une demarche de planification integree des ressources permettant d'ameliorer l'efficacite globale du plan de transport, d'accroitre la competitivite de la sncf et d'ameliorer la qualite de ses services. nous avons propose une formalisation du probleme de planification integree engins/adc et des algorithmes performants qui s'appuient sur une approche par relaxation lagrangienne pour resoudre de maniere efficace la problematique etudiee. cette approche repose sur l'exploitation de deux briques logicielles developpees a la sncf pour resoudre chacun des sous-problemes de planification des engins et des adc. les algorithmes ont ete testes experimentalement avec des donnees reelles de la region ter bretagne. differentes evolutions des modeles et des algorithmes ont ete etudiees pour rendre ces derniers plus efficaces. les tests de validation sur des jeux de donnees reelles a une echelle industrielle sont encourageants et montrent des gains potentiels allant jusqu'a 4% des adc exploites par rapport a une approche traditionnelle (sequentielle). / The planning of railway production at the french national railways (sncf) is currently based on a mainly sequential process in which the design of railway timetabling widely conditioning design planning of railway equipment (rolling stock), then one of the train drivers (driver rosters). this strategy of sequential planning of railway resources massively adopted for practical and scientific reasons (expertise, complexity of the railway system, etc.). however, this strategy generates solutions which can be more expensive and less robust to uncertainties, because decisions taken at any given stage can significantly reduce the overall feasible solutions of the following steps.given this situation and the strong interaction between these heterogeneous and very expensive resources, the thesis deals with the feasibility and inputs of a process where these critical resources could be planned and optimized in an integrated way. the thesis focuses on the feasibility study, prototyping and validation of an integrated approach for planning rolling stocks and drivers, so as to improve the efficiency of the overall transportation plan, increase sncf competitiveness and enhance the quality of its services. we propose a mixed integer linear programming formulation of the rolling stock/ train drivers integrated planning problem. in this mathematical model, each planning sub-problem is formalized and coupling constraints are further introduced to model the interdependencies of these two resources when they are simultaneously used for train production. in this heuristic, the solution of the lagrangian dual and the calculation of feasible solutions are performed by calling two proprietary software modules available at sncf for planning rolling stocks and train drivers. the heuristic is tested experimentally with real data from the ter bretagne region, and several evolutions are introduced in the models and algorithms so as to improve their performances.validation tests on of real data sets at an industrial scale are encouraging and, when compared to a traditional (sequential) approach, show gain of up to 4% for train drivers used.