Détection continue de collisions entre triangles et fonctions distance signée en temps réel

Abstract

From virtual training for surgeons and pilots to live obstacle avoidance and trajectory optimization in aerospace and robotics, many modern systems rely on real-time physics simulations. Accurate and robust collision detection is key for useful simulations. Yet, it is excessively difficult to achieve in real-time applications. The proper modelling of objects directly affects the realism of physics-based simulations and how efficiently we can simulate physical phenomena. This thesis introduces efficient solutions for high-quality collision detection between triangle meshes and signed distance functions (SDFs). The continuous collision detection (CCD) problem is tackled with a novel spatio-temporal optimization to seek the exact time of impact (TOI) between a triangle and an SDF isosurface. This method offers improved robustness over existing point sampling methods, and outperforms recent triangle-SDF discrete collision detection (DCD) algorithms. Also proposed, is a novel adaptive refinement method to efficiently provide supplemental contact information over large triangles. Efficient coarse geometry can thus be used to reduce the cost of collision detection involving triangle meshes. A comparison against state-of-the-art algorithms is provided, and exposes the capabilities and advantages of the proposed approach through simulating thousands of difficult collision scenarios. Experimental results demonstrate that this work improves the quality of physics simulations while maintaining real-time performance. Overall, this work provides the first working method for robust triangle SDF CCD, and opens the door to a wide array of new potential applications for physics-based simulations where small and fast moving objects interact.

Date	14 Dec 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Sheldon Andrews (Supervisor)