基于深度学习的低成本堆叠物料定位系统

A Low-Cost Location System Based on Deep Learning for Stacked Materials Sorting

针对堆叠物料无序分拣中传统定位方法硬件成本高、检测精度低等问题,设计了一种基于深度学习的堆叠物料定位系统.以单目光学相机采集得到的图像作为输入数据,利用单阶段检测算法得到候选目标,采用卷积神经网络进行目标筛选,最后对筛选后的目标感兴趣区域图像进行特征点回归,得到目标的类别、坐标和角度.堆叠物料定位系统由于无需昂贵的深度相机,且算法的鲁棒性较高,降低了硬件成本,提高了检测精度.在真实场景的测试结果显示,新系统的定位误差降低到了0.3 cm以内.

To improve industrial production efficiency,a low-cost location system based on deep learning was proposed for stacked materials sorting.Firstly,taking the images got from monocular optical cameras as the input data,a one-stage detection method was used to obtain the candidate objects.Then,a deep convolution neural network was used to classify the adjacent ROIs of objects to filter the possible candidates.Finally,some interest images of the filtered candidates were processed to get the key shape and location of the object.Since the robustness of the algorithm and without the use of the expensive depth camera,the new method can reduce hardware cost and improve detection accuracy.The test results in the real scene show that the location error of the new system can be reduced to less than 0.3 cm.