Description du sujet de thèse

Athletic performance today heavily relies on data analysis from both training and competition. In many sports, especially "machine sports" like cycling, athletes are equipped with various sensors, making it easy to access quantities such as position, speed, and power output. However, for a coach or a national team, this information is obviously limited to their own athletes, as opponents' data is not shared during competitions.
In contrast, video tracking—widely used in many sports—provides information regardless of the athlete, since both opponents and team members can be filmed. However, video tracking in track cycling is not yet widespread. This is mainly due to the large and geometrically complex (non-flat and not known a priori) nature of the track, making camera calibration difficult.
The world's first video tracking system for cyclists was launched at the National Velodrome in Saint-Quentin-en-Yvelines (SQY) in October 2022. Other velodromes plan to adopt similar systems, but setting up the hardware infrastructure and software development takes several months. Currently, such data is not available at other velodromes. Moreover, during the 2024 Olympic Games, the Paris 2024 Organizing Committee took full control of the velodrome, requiring the camera system to be shut down, preventing access to the data it previously generated.
These considerations led the French Cycling Federation (FFC), as part of the ANR-funded project THPCA (High Performance in Cycling and Rowing – 2021-2024), to request the development of a portable detection system. In 2023–2024, we developed a Matlab code that automatically and very precisely detects the positions of cyclists during track competitions, using only two fixed-camera videos.
This system has already been successfully used to analyze the performances of both French and international athletes during the Paris 2024 Olympics. Since key metrics like speed and acceleration are essential for such analyses, the cyclist's position over time must be determined with high accuracy.
The PhD project aims to improve the tracking code, involve FFC staff in its usage, and use it to model certain physical and tactical parameters in races.
To begin, the PhD candidate will learn to operate this complex code to produce cyclist trajectories. The key steps include:
• Image stabilization
• Image calibration relative to the velodrome and determination of its geometry
• Automatic cyclist detection in images
• Precise localization of cyclists on the track
• Identification of individual cyclist trajectories
Each step will require in-depth understanding of the underlying algorithms. Notably, the current detection algorithm does not use Artificial Intelligence (AI), which means it requires no training phase. However, given recent advances in AI and its natural application in image analysis, it will be of interest to compare detection performances with and without AI.
Moreover, our current method cannot analyze mass-start races due to the large number of cyclists (about twenty), but AI might allow us to address this case. The first step of the PhD will be to propose an AI-based detection model and to conduct a detailed performance comparison with the non-AI model.
Once the most effective system (AI or non-AI) is finalized and mastered by the PhD student, a tool will be developed to compute the power output of athletes—one of the most used metrics by cycling coaches. Though typically measured by onboard sensors, a precise trajectory combined with a detailed energy model over the course of a race can also allow non-instrumented estimation of an athlete's power over time, making it possible to evaluate opponents as well.
With this data (trajectory, speed, power), several analyses will be conducted:
• Drafting effect analysis: Determine how the drag reduction behind another cyclist depends on both longitudinal and lateral distance. Though primarily studied in wind tunnels, real-race fluctuations may produce different effects. We recently developed a project using onboard distance sensors, but they cannot be used in competition. Estimating inter-cyclist distances via video, and correlating with power data, will guide tactical decisions in team events (e.g., rider order in team pursuit, gear selection in team sprint). New video sessions with French national athletes will be organized to expand the existing dataset.
• Race strategy analysis: For example, in 200 m flying sprints or team sprints, knowing the power evolution via video allows comparison of tactical decisions—e.g., when to start the dive in a 200 m, or when to initiate a teammate overtake in team sprint.
• Drag area measurement during competition: Drag areas are usually measured in wind tunnels. However, knowing both power (from onboard sensors) and trajectory (from video) allows direct drag area estimation during races. This could reveal suboptimal rider positions or aerodynamic evolution throughout a race. Wind tunnel sessions may be conducted in collaboration with the Aerotechnical Institute wind tunnel (partner of THPCA), located near SQY.

Contexte de travail

The project is built on solid foundations from the ANR THPCA project, which developed both scientific knowledge and close human collaboration with FFC staff. This explains the co-supervision with Iris Sachet, FFC sport scientist for the past four years. Her involvement ensures connection with real-world concerns of athletes and staff. Emmanuel Brunet, scientific advisor to the FFC and our main THPCA contact, has also expressed interest in contributing. For the AI component, we will benefit from the expertise of Julian Tachella (CNRS researcher at ENS de Lyon), a specialist in AI for image analysis. The core tracking code already exists and performs well, enabling the PhD student to start work immediately.

Contraintes et risques

Confidentiality regarding foreign sports federations.