基于深度神经网络复杂场景下的机器人拣选方法

针对提高快递包裹的分拣效率和识别准确率，提出了一种基于深度神经网络复杂场景下的机器人拣选方法.首先，提出一种改进的目标检测算法，通过将多层浅层特征图与最终的特征图进行融合，提取更加细节的特征，以提升识别的速度与精度；其次，提出了一种基于关键点的级联卷积最优拣选位置检测网络模型，对包裹最优拣选位置进行实时预测估计；最后，结合目标包裹最优拣选框与场景的深度信息和基于三维信息的目标姿态估计算法实现机器人拣选，并通过实验验证了该方法的有效性.

Robotic Sorting Method in Complex Scene Based on Deep Neural Network

A robotic sorting method based on deep neural network in complex scene is proposed to improve the sorting efficiency and recognition accuracy of parcels. The sorting method consists of three main parts. Firstly, the improved object detection algorithm is proposed. More detailed features are extracted by combining the multi-layer shallow layer with the final feature map to improve the speed and accuracy of recognition. Then, an optimal grab position detection network based on cascading convolution of key points is proposed to realize real-time estimation of the optimal sorting position of the parcel. Finally, by combining with the target capture optimal frame and the depth information of the scene, the robotic sorting operation can be completed by the target pose estimation algorithm based on the three-dimensional information, and the effectiveness of the method is verified by experiments.