Binocular localization method for pear-picking robots based on YOLO-CDS and RSIQR module collaborative optimization

Abstract

Variations in lighting conditions and obstructions significantly impact the localization accuracy of pear-picking robots in standardized orchards. This study proposed a binocular localization method that integrated an improved You Only Look Once (YOLO-CDS) detection network with a Recurrent All-Pairs Field Transformers for Stereo and Interquartile Range statistical optimization module (RSIQR). By constructing a dual-module collaborative architecture of “detection-depth value optimization,” the method systematically addressed challenges such as lighting, obstructions, and background interference. In the detection module, the YOLOv8-based YOLO-CDS model introduced the C2f-Efficient Multi-scale Bottleneck Convolution module (C2f-EMBC), DySample module, and Shape Intersection over Union (Shape-IoU) to enhance feature extraction and bounding box accuracy. In the RSIQR module, the Recurrent All-Pairs Field Transformers for Stereo (RAFT-Stereo) algorithm replaced the Semi-Global Matching (SGM) algorithm in the binocular camera system to improve disparity estimation. Finally, the depth values from YOLO-CDS and RAFT-Stereo were optimized using the Interquartile Range (IQR) method to obtain more accurate depth information. Experimental results showed that the YOLO-CDS model achieved mAP@50 of 97.5 % on the multi-class pear dataset, an improvement of 1.9 percentage points over previous models. Additionally, the maximum localization error of pear coordinates under daytime front lighting, daytime backlighting, and nighttime supplementary lighting conditions within 300 mm to 1000 mm was 2.357 % ± 0.415 %. Orchard picking tests showed that for unobstructed targets, the proposed method achieved a localization success rate of 100 % and a picking success rate of 91.93 %. These results show that the proposed method, combined with a tailored picking strategy, reduces invalid actions while ensuring hardware safety and continuous operation, thereby contributing to improved picking efficiency. This study provides robust technical support for the application of pear-picking robots in standardized orchard environments.