Visual detection of moving stacked objects based on efficient multi-scale grouping and improved multi-head self-attention

To address the issue of fruit stacking and obstructing target fruits during daily fruit sorting, this paper presents a novel object detection algorithm that leverages efficient multi-scale grouping and enhanced multi-head self-attention. The proposed target detection algorithm is based on real-time detection transformer (RT-DETR) as the baseline to simplify optimization challenges and enhance robustness. Additionally, we introduce efficient multi-scale attention to preserve channel information, optimize the multi-head self-attention, and adopt cascade grouping to reduce computational redundancy. Furthermore, we use a new loss function (Inner-MPDIoU) combined with a bounding box similarity comparison metric (MPDIoU) and inner idea to enhance the accuracy of detecting moving occluded targets. Experimental results demonstrate that the optimized RT-DETR algorithm achieves an average accuracy of 96.3% in detecting moving stacked fruit models with a detection speed of up to 67 FPS. This confirms the effectiveness of our algorithm in matching and recognizing blocked fruit targets, surpassing common algorithms for recognizing obstructed targets.