Three-dimensional (3D) real-time object detection and tracking is an important task in the case of autonomous vehicles and road and railway smart mobility, in order to allow them to analyze their environment for navigation and obstacle avoidance purposes. In this paper, we improve the efficiency of 3D monocular object detection by using dataset combination and knowledge distillation, and by creating a lightweight model. Firstly, we combine real and synthetic datasets to increase the diversity and richness of the training data. Then, we use knowledge distillation to transfer the knowledge from a large, pre-trained model to a smaller, lightweight model. Finally, we create a lightweight model by selecting the combinations of width, depth & resolution in order to reach a target complexity and computation time. Our experiments showed that using each method improves either the accuracy or the efficiency of our model with no significant drawbacks. Using all these approaches is especially useful for resource-constrained environments, such as self-driving cars and railway systems.

Monocular 3D object detection (Mono3D) is essential for autonomous driving and augmented reality, yet its performance degrades significantly at night due to the scarcity of annotated nighttime data. In this paper, we investigate the use of style transfer for nighttime data augmentation and evaluate its effect on individual components of 3D detection. Using CycleGAN, we generated synthetic night images from daytime scenes in the nuScenes dataset and trained a modular Mono3D detector under different configurations. Our results show that training solely on style-transferred images improves certain metrics, such as AP@0.95 (from 0.0299 to 0.0778, a 160% increase) and depth error (11% reduction), compared to daytime-only baselines. However, performance on orientation and dimension estimation deteriorates. When real nighttime data is included, style transfer provides complementary benefits: for cars, depth error decreases from 0.0414 to 0.021, and AP@0.95 remains stable at 0.66; for pedestrians, AP@0.95 improves by 13% (0.297 to 0.336) with a 35% reduction in depth error. Cyclist detection remains unreliable due to limited samples. We conclude that style transfer cannot replace authentic nighttime data, but when combined with it, it reduces false positives and improves depth estimation, leading to more robust detection under low-light conditions. This study highlights both the potential and the limitations of style transfer for augmenting Mono3D training, and it points to future research on more advanced generative models and broader object categories.

Cette thèse présente plusieurs contributions au développement de la détection monoculaire d'objets en 3D pour des applications autonomes de monorail. Tout d'abord, une plateforme expérimentale de monorail a été conçue et mise en œuvre, ainsi que des lois de contrôle pour assurer sa stabilisation. Sur cette base, des améliorations ont été apportées aux systèmes de détection monoculaire d'objets en 3D en mettant en œuvre les améliorations existantes pour les méthodes de détection d'objets en 2D. Pour améliorer encore les performances de détection, de nouvelles fonctions de transition de perte ont été développées et évaluées, permettant de donner la priorité à des métriques spécifiques pendant l'apprentissage du modèle. En outre, des méthodes ont été proposées pour combiner efficacement des ensembles de données réelles et synthétiques, en relevant des défis tels que les incohérences dans les matrices intrinsèques des caméras, la normalisation des ensembles de données et l'augmentation des données. Ces approches visent à améliorer les capacités d'apprentissage et de généralisation des modèles. Enfin, des techniques de transfert de style ont été employées pour augmenter les ensembles de données de formation en simulant diverses conditions environnementales, telles que les changements d'éclairage et de temps.