关键词: 苹果智能采摘, 双目视觉, YOLOv3, 深度信息, 算法

Abstract: The accurate recognition and three-dimensional positioning of apple picking targets in natural scenes are the key technologies of apple intelligent picking equipment. Integrating YOLOv3 algorithm and binocular vision technology, the samples in a variety of natural scenes were trained through YOLOv3 algorithm, the recognition model was constructed, apple imageswere obtained by binocular vision, the two-dimensional coordinates of the target apple in the image were obtained by using the YOLOv3 model, and then the depth coordinate information was obtained by using the binocular vision parallax principle to realize the three-dimensional spatial positioning of the target apple. The algorithm was applied to apple recognition and location in different natural scenes, and the recognition effect and location accuracy were evaluated. The results showed that the minimum relative error was 0.193% and the maximum relative error was 3.670% under the condition of uneven illumination, shadow on the fruit and mutual occlusion; In the case of insufficient illumination and mutual occlusion at night, the minimum relative error was 0.831% and the maximum relative error was 4.417%; The minimum relative error was 0.176% and the maximum relative error was 4.205% when dew forms reflection on the surface of apple and the fruits block each other; When the light was weak, the shadow was small and there was occlusion, the minimum relative error was 0.168%, and the maximum relative error was 3.776%. The algorithm used in the study only needed an appropriate number of samples to meet the recognition and positioning training in different scenes, and mAP （mean Average Precision） in different scenes was 96.60%. The algorithm has strong stability, can identify the target apples in different application scenarios such as overlap, night and light refraction, and can better meet the accuracy requirements of intelligent robot in recognition and positioning.

Key words: apple intelligent picking, binocular vision, YOLOv3, depth information, algorithm