Abstract: In order to realize automatic picking of passion fruit under the environment of field shelving, YOLOv3 deep convolutional neural network was used to establish the fruit location model of passion fruit under the complex background. In this method, a single convolutional neural network was used to traverse the whole image and return to the category and location of the target thus to realize the direct end-to-end target detection. The training set used 400 manually labeled images and the test set used 100 images, which included a total of 3,071 passion fruit samples. The average accuracy mean mAP of the trained model under the test set was 97.66%. When the confidence threshold of 0.65 was used, the accuracy rate was 98%, the recall rate was 94%, and the crossover ratio of IOU was 83.96%.