Robotic harvesting shows a promising aspect in future development of agricultural industry. However, there are many challenges which are still presented in the development of a fully functional robotic harvesting system. Vision is one of the most important keys among these challenges. Traditional vision methods always suffer from defects in accuracy, robustness, and efficiency in real implementation environments. In this work, a fully deep learning-based vision method for autonomous apple harvesting is developed and evaluated. The developed method includes a light-weight one-stage detection and segmentation network for fruit recognition and a PointNet to process the point clouds and estimate a proper approach pose for each fruit before grasping. Fruit recognition network takes raw inputs from RGB-D camera and performs fruit detection and instance segmentation on RGB images. The PointNet grasping network combines depth information and results from the fruit recognition as input and outputs the approach pose of each fruit for robotic arm execution. The developed vision method is evaluated on RGB-D image data which are collected from both laboratory and orchard environments. Robotic harvesting experiments in both indoor and outdoor conditions are also included to validate the performance of the developed harvesting system. Experimental results show that the developed vision method can perform highly efficient and accurate to guide robotic harvesting. Overall, the developed robotic harvesting system achieves 0.8 on harvesting success rate and cycle time is 6.5 s.

Autonomous harvesting shows a promising prospect in the future development of the agriculture industry, while the vision system is one of the most challenging components in the autonomous harvesting technologies. This work proposes a multi-function network to perform the real-time detection and semantic segmentation of apples and branches in orchard environments by using the visual sensor. The developed detection and segmentation network utilises the atrous spatial pyramid pooling and the gate feature pyramid network to enhance feature extraction ability of the network. To improve the real-time computation performance of the network model, a lightweight backbone network based on the residual network architecture is developed. From the experimental results, the detection and segmentation network with ResNet-101 backbone outperformed on the detection and segmentation tasks, achieving an $F_1$ score of 0.832 on the detection of apples and 87.6% and 77.2% on the semantic segmentation of apples and branches, respectively. The network model with lightweight backbone showed the best computation efficiency in the results. It achieved an $F_1$ score of 0.827 on the detection of apples and 86.5% and 75.7% on the segmentation of apples and branches, respectively. The weights size and computation time of the network model with lightweight backbone were 12.8 M and 32 ms, respectively. The experimental results show that the detection and segmentation network can effectively perform the real-time detection and segmentation of apples and branches in orchards.

Visual perception and modelling are essential tasks in many vision-dependent robotic tasks. This work presents a robotic vision system to perform fruit recognition, modelling, and environment modelling for autonomous apple harvesting. The fruit recognition applies a deep-learning model Dasnet to perform detection and segmentation on fruits, and segmentation on branches. Fruit modelling localises the centre and computes the grasp pose of each fruit based on Hough Transform. Environment modelling adopts Octrees to represent the occupied space within the working environment of the robot. The robot control computes the path and guide manipulator to pick the fruits based on the computed 3D model of the crop. The developed method is tested in both laboratory and orchard environments. Test results show that fruit recognition and modelling algorithm can accurately localise the fruits and compute the grasp pose in various situations. The Dasnet achieves 0.871 on F1 score of the fruit detection. Fruit modelling achieves 0.955 and 0.923 on the accuracy of the fruit centre estimation and grasp orientation, respectively. To illustrate the efficiency of the vision system in autonomous harvesting, a robotic harvesting experiment by using industry robotic arm in a controlled environment is conducted. Experimental results show that the proposed visual perception and modelling can efficiently guide the robotic arm to perform the detachment and success rate of harvesting is improved compared to the method which does not compute the grasp pose of fruits.

Camellia oleifera fruits are randomly distributed in an orchard, and the fruits are easily blocked or covered by leaves. In addition, the colors of leaves and fruits are alike, and flowers and fruits grow at the same time, presenting many ambiguities. The large shock force will cause flowers to fall and affect the yield. As a result, accurate positioning becomes a difficult problem for robot picking. Therefore, studying target recognition and localization of Camellia oleifera fruits in complex environments has many difficulties. In this paper, a fusion method of deep learning based on visual perception and image processing is proposed to adaptively and actively locate fruit recognition and picking points for Camellia oleifera fruits. First, to adapt to the target classification and recognition of complex scenes in the field, the parameters of the You Only Live Once v7 (YOLOv7) model were optimized and selected to achieve Camellia oleifera fruits’ detection and determine the center point of the fruit recognition frame. Then, image processing and a geometric algorithm are used to process the image, segment, and determine the morphology of the fruit, extract the centroid of the outline of Camellia oleifera fruit, and then analyze the position deviation of its centroid point and the center point in the YOLO recognition frame. The frontlighting, backlight, partial occlusion, and other test conditions for the perceptual recognition processing were validated with several experiments. The results demonstrate that the precision of YOLOv7 is close to that of YOLOv5s, and the mean average precision of YOLOv7 is higher than that of YOLOv5s. For some occluded Camellia oleifera fruits, the YOLOv7 algorithm is better than the YOLOv5s algorithm, which improves the detection accuracy of Camellia oleifera fruits. The contour of Camellia oleifera fruits can be extracted entirely via image processing. The average position deviation between the centroid point of the image extraction and the center point of the YOLO recognition frame is 2.86 pixels; thus, the center point of the YOLO recognition frame is approximately considered to be consistent with the centroid point of the image extraction.