Performance evaluation of newly released cameras for fruit detection and localization in complex kiwifruit orchard environments

• DOI: 10.1002/rob.22297
• Publication Date: 2024-01-30T11:15:03Z
• Authors: Xiaojuan Liu; Xudong Jing; Hanhui Jiang; Shoaib Younas; Ruiyang Wei; Haojie Dang; Zhenchao Wu; Longsheng Fu
• Keywords: Calyx; RGB color model; Orchard

Abstract Consumer RGB‐D and binocular stereo cameras were applied to fruit detection localization. However, few studies are documented on performance comparison of newly released under same scene in complex orchard. This study evaluates consumer based YOLOv5x for kiwifruit localization selection optimal one with better application orchard environment. Firstly, Azure Kinect, RealSense D435, ZED 2i employed capture images canopies. Subsequently, was train detect kiwifruits calyxes the images. Meanwhile, an overlap‐partitioning strategy calyx detection. Additionally, spatial coordinate transformation performed by integrating camera's extrinsic parameters depth map generated each camera. Finally, three‐dimensional coordinates calculated compared ground truth, followed accuracy analyzed. Results show that obtained mean average precision 93.2%, 91.3%, 95.8% three detection, respectively. Overlap‐partitioning improved significantly increased 13.00%, 16.30%, 7.70%, The absolute deviation Y‐axis is relatively high at 8.44 mm 6.67 while D435 achieved minimum 10.42 X‐axis 18.33 Z‐axis. Average speed image 0.164 s, 0.037 0.062 s 2i, These results indicate excellent than Kinect orchard, which could be a valuable reference other orchards select camera capacity.