Edge AI for Monocular Cone Perception in Formula Student Driverless Racing

Reliable perception is a fundamental requirement for autonomous racing systems, where real-time operation and limited onboard computational resources impose strict constraints on algorithm design and deployment. In the context of Formula Student Driverless (FSD), cone detection and localization are key perception tasks for track understanding and downstream planning. This paper presents an embedded monocular cone perception pipeline for autonomous racing, designed as a low-cost and low-complexity solution for real-world deployment. The proposed system combines a calibrated monocular camera, a YOLO-based object detector, and a geometric localization module based on the pinhole camera model to estimate cone positions in the vehicle reference frame. The pipeline is deployed on an NVIDIA Jetson Orin Nano platform and evaluated in terms of detection accuracy, localization error, runtime performance, and power consumption. Experimental results show that the best performing configuration, based on INT8-quantized YOLOv10s, achieves an mAP@50:95 of 48.53% and a localization RMSE of 0.760 m, while sustaining 29.90 FPS with an end-to-end power consumption of approximately 7.1 W. A lighter YOLOv10n configuration achieves 46.73% mAP@50:95, 0.840 m RMSE, and 31.80 FPS at 6.8 W, providing a favorable trade-off between accuracy and efficiency. These results indicate that monocular vision can provide an effective embedded perception baseline for FSD autonomous racing under tight cost, power, and complexity real-time constraints. The proposed pipeline is conceived as a perception building block for onboard artificial intelligence (AI) in autonomous racing systems.