Research On Target Recognition Technology Of Jujube Picking Robot

The acreage of jujube planting flourishing with the development of China’s social economy and the national devotion to the cultivation.The traditional manual picking,however,is still a prevalent method,which remains labor intensiveness,low efficiency and high cost.Although mechanical picking is available,apparently its high loss rate dissatisfies requirements of modern agriculture.The development of intelligent manufacturing provides more efficient feasibility to overcome difficulties above.Aiming at the inefficiency of manual picking and large loss of mechanical picking,development of deep-learning based target recognition technology is proposed for jujube picking robot in this paper,it provides data support and solution for the automatic picking of jujube fruit.The following research are carried out:1.Establishment of jujube fruit labeling database.It is necessary to set up jujube dataset due to the insufficiency of open dataset.In this paper,images of both singled and clustered jujube fruits under different conditions are acquired,filtered,and enhanced,and then a database of jujube images is established.The images in dataset were labeled by Labelling to form a dataset of 13659 jujube fruit pictures.2.Lightweight target recognition network build up.It is tremendous load to migrate the network and the recognition speed latency for embedded devices since large number of parameters of yolov3 target recognition algorithm model.In this paper,a Darknet-53 backbone network in yolov3 is modified into a lightweight MobileNetV3 network,which reduces the model parameters down to 1/7 of the original network.The modified MobileNetV3-yolov3 network is also improved,which increases the prediction of large feature map size,enhances the recognition accuracy of jujube fruit,optimizes the learning rate and batch size,and obtaining the best verification effect is achieved by selecting the strategy of prior box.The average accuracy of the improved network is 88.24%.Compared with the yolov3 algorithm,the speed is almost doubled,reaching of 34 frames per second.Therefore,it is possible to deploy the algorithm in real embedded devices.3.Obta7in the depth data of jujube fruit and construct the target recognition system of jujube picking robot.The picking robot is unable to grasp the jujube fruit by using only pixel coordinates in the actual work,therefore three-dimensional coordinates of jujube fruit is indispensable for the grasping work.In this paper,the Intel RealSense D435i depth camera is adopted in shooting jujube fruits.By using both external and internal parameters of the camera and the conversion of pixel coordinates into three-dimensional space coordinates,the depth in the real space of jujube fruits and accurate positioning are acquired.In order to reveal the target recognition process on jujube picking robot more figuratively,the target recognition system on the jujube picking robot is constructed,the display of recognition results and three-dimensional coordinates in the process of target recognition on the jujube picking robot are accomplished.