THE AUTONOMOUS SYSTEM
We operate our car with an autonomous system developed completely in-house by our software divisions, utilizing LiDAR, cameras, image recognition softwares and a whole lot more.
The perception system processes data from LiDAR and cameras to detect and classify cones. LiDAR point clouds undergo geometric filtering, ground removal, and clustering (DBSCAN) to isolate cone-like shapes.
Camera images are analyzed with YOLOv8/YOLOv10 to classify cone colors. Sensor fusion combines LiDAR’s precise positioning with camera-based color recognition, ensuring accurate environmental mapping. Synchronization between sensors is achieved through trigger-based calibration.
The SLAM system localizes the vehicle and constructs a track map in real time. GraphSLAM models the environment as a graph where nodes represent vehicle poses and cones, and edges define spatial constraints.
Optimization minimizes localization errors by refining the vehicle’s estimated trajectory. This approach enables precise positioning despite sensor noise and dynamic conditions, ensuring reliable autonomous navigation.
Path planning generates an optimal trajectory using geometric and probabilistic techniques. Delaunay Triangulation extracts the track centerline by connecting cones of different colors and identifying midpoints between them. Trajectory optimization refines the route by minimizing curvature and considering constraints such as vehicle dynamics and track boundaries. A velocity profiler then assigns a feasible speed at each point along the trajectory, based on curvature and acceleration limits. This process ensures a safe, efficient, and dynamically feasible path for the autonomous car.
Path tracking executes the planned trajectory with high precision using a combination of Pure Pursuit and Model Predictive Contouring Control (MPCC), depending on track conditions.
We adopt two path tracking strategies:
– MPCC is used for high-speed segments, where optimal control and dynamic adaptation are crucial.
– Pure Pursuit with a velocity profiler is chosen for its reliability and computational efficiency, especially in conditions where robustness is prioritized.
MPCC predicts the vehicle’s future behavior and optimizes steering and speed to maximize progress along the path, integrating a dynamic vehicle model to ensure smooth navigation and high-speed stability. In contrast, Pure Pursuit ensures responsive and efficient tracking with a simpler implementation, making it ideal for more predictable or constrained scenarios.
