Enhancing Environmental Perception in Formula Student: A Robust LiDAR Pipeline for Real-Time Cone Detection.

This thesis investigates the performance enhancements achieved by integrating a high density LiDAR sensor into the autonomous perception stack of a Formula SAE (FSAE) Driverless vehicle. Developed within the Unibo Motorsport team, the research focuses on the transition from a legacy 16 layer SICK MultiScan165 to a state of the art 128 layer Hesai OT128 LiDAR. The core contribution of this work is the redesign of the environmental perception pipeline, shifting from a sparse 2.5D layer-wise heuristic approach to a robust volumetric 3D processing architecture. The new pipeline implements advanced stages including Patchwork++ for resilient ground segmentation on uneven racing terrains, Voxel Grid Downsampling to maintain real time processing of over 3.4 million points per second, and Euclidean Cluster Extraction combined with compactness filtering for precise traffic cone identification. Experimental results demonstrate a five fold increase in the vehicle’s effective path horizon, expanding from 4 meters with the SICK sensor to 25 meters with the OT128. At a nominal speed of 30 km/h, this extension provides the trajectory planner with a 3.0 second temporal window, representing a significant improvement over the unstable 0.48 second window afforded by the legacy system. While the high-density sensor increases computational and electrical demands, the findings conclude that the superior spatial accuracy and look ahead capability are essential for maintaining safety and stability in high speed autonomous racing. To further evolve the system, future work will focus on implementing tracking by detection techniques, such as AB3DMOT, to mitigate sensor noise and ”flickering,” alongside exploring deep learning architectures like IA-SSD and Point-Voxel CNN to enhance semantic understanding and object classification.