Embedded Vision System using Raspberry Pi for Ripe Strawberry Detection and Localization using YOLOv5s-CGhostnet

초록

Automated harvesting of delicate fruits, such as strawberries, requires vision systems that provide high accuracy and a rapid response. Implementing these systems on embedded platforms imposes strict constraints on power consumption, memory usage, and computational resources. Conventional deep learning models typically fail to meet the stringent memory and computational limitations of such devices. Building upon the previously designed lightweight YOLOv5s-CGhostNet architecture, this study presents its conception, adaptation, and implementation for real-Time embedded operation on a Raspberry Pi. The system integrates post-Training quantization to reduce model size and enhance inference speed, enabling efficient performance under edge conditions. Beyond fruit detection, the proposed framework incorporates the precise localization of peduncle cutting points on ripe strawberries through feature engineering techniques applied after detection. Experimental results demonstrate that the optimized YOLOv5s-CGhostNet model for Raspberry Pi, achieved via INT8 post-Training quantization with TensorFlow Lite, yields a mAP of 87.6% and an inference speed of 13 FPS. Post-detection feature engineering localizes peduncle cutting points (2.1 px 2D error) and fuses it with RealSense D455 depth, achieving a 4.2 mm 3D world error in a range of 20-80 cm, with 92% accuracy within a 5mm tolerance, enabling real-Time embedded applications in autonomous strawberry harvesting. © 2026 IEEE.

키워드