ToF 3D Vision Detection and Localization of Soft Packaging Bags Based on Deep Learning

The detection of soft packaging bags is a key step in the process of soft packaging bags unpacking. Time-of-flight(ToF) camera is used as vision sensor and a multi-scale detection method based on deep learning is proposed to solve the deformation and shielding problems of soft packaging bags. This method is improved on the basis of YOLOv3. Aiming at the texture disorder and shape change caused by the deformation of the soft packaging bags, the deformation convolution is used to replace the standard convolution for feature extraction. In view of the slow detection speed of YOLOv3, the inverted residual module based on depth separable convolution is used to replace the residual module. Aiming at the problem of non-detection caused by the occlusion between soft packaging bags, the loss function in YOLOv3 is improved and the size of anchor box is adjusted. The test is carried out when there were different degrees of shielding between the soft packaging bags and different degrees of deformation of the soft packaging bags. The experimental results show that the detection speed of this method reaches 0.8s/frame, the recall rate reaches 99%, and the relative positioning accuracy reaches 2 cm.