Real-Time Recognition and Localization of Kiwifruit Based on Improved YOLOv5s Algorithm

To address the recognition and localization challenges posed by occlusion and overlap due to dense fruit growth as well as differences in ripeness, size, morphology, and color in smart Kiwifruit harvesting. This study proposes a real-time Kiwifruit recognition and localization method based on an enhanced YOLOv5s algorithm combined with binocular vision technology. First, data enhancement is implemented in the training strategy to ensure the model’s efficient learning and generalization ability in complex environments; then the Coordinate Attention mechanism is introduced to highlight the key features of the image and improve detail detection; Secondly, the Bidirectional Feature Pyramid Network structure is used to optimize the feature fusion and strengthen the information exchange between different layers through bidirectional connectivity; then, the loss function is optimized to improve the bounding box localization accuracy; and finally, the combined with the binocular vision stereo matching algorithm Semi-Global Block Matching to obtain the spatial location information of Kiwifruit. The experimental results show that the improved YOLOV5s improve the detection accuracy by 6.1%, recall by 5%, and average precision mean by 7.3% compared with its predecessor model. The average recognition of a single image takes only 0.125 seconds. The error of the target’s actual world coordinates and the localization coordinates is within 0.5 cm, mainly concentrated within 0.3 cm, showing the method’s reliability and stability to meet the demand for intelligent Kiwifruit picking.