Realization Of Pallet Picking Scheme Based On Monocular Vision Guidance

There are a large number of cargo handling in the fields of manufacturing and logistics.Pallet stacking is the most important storage method.Industrial handling vehicles that pick up pallets are called forklifts.Forklifts need to accurately pick up the stacked pallets,safely transport them to the designated location,and finally unload them smoothly.At present,the main devices of automatic guided forklift are laser and radio frequency sensing elements with high precision but expensive.This paper chooses a monocular visible light camera to design and implement a pallet picking method that meets precise requirements,real-time judgment,high robustness,low cost,and strong scalability.The main contents include:(1)According to the characteristics of unmanned vehicle operation,inspect pallets,goods and workers to obtain the approximate location of the target.Use visible light camera to collect unmanned vehicle operation scenes,and make data sets with pallets,goods and workers as labels.Based on the YOLOv4 deep learning target detection algorithm,the network model is trained and then tested.The accuracy is high,and the processing time for each frame is short,which can meet real-time requirements.(2)Based on the Siammask network to track the pallet during the picking process,adjust the Youtube-VOS-2019 public data set for the training of the single-target tracking network in this paper,and add the pallet data sample of this paper on the basis of this data set.Determine the image area and approximate pose of the target pallet according to the mask output from the network.(3)Regarding the problem that the Siammask network is interfered by non-tracking target pallets of the same shape,color,and orientation,according to the characteristics of this article,combined with the known driving movement or rotation speed,the conversion between world coordinates and image coordinates is established,and it is estimated that The position of the target pallet in the image of the current frame,and the intercepted image is used as the input of the Siammask network.And introduce the position weight to modify the Siammask network position score judgment,which greatly reduces the possibility of predicting non-tracking targets to interfere with the pallet,and finally solves the problem.(4)The mask output by the Siammask network is further processed to obtain the pallet pixels.After expansion and Gaussian filtering,the contour is extracted through the Canny operator.The Hough transform obtains the candidate straight lines of the straight edge contour of the top surface of the pallet,and reduces the angle traversal range of the Hough transform according to the slope of the straight edge of the image target in the previous frame or the slope of the minimum circumscribed rectangle of the current mask.Combining the characteristics of the straight edge contour pixels of the pallet in this paper,clustering and screening are carried out,and the upper surface edge of the target pallet in the image frame is obtained,which is used for driving judgment.(5)The experimental car uses the Raspberry Pi 4B as the main board,installs Mecanum wheels,and encapsulates the movement and rotation instructions of the experimental car.The experimental car is the client,the personal computer is the server,and the TCP connection is established.The camera image of the experimental vehicle is transmitted to the personal computer in real time,the personal computer processes the image to make a judgment,and feeds back the driving instructions of the experimental vehicle in real time,and the experimental vehicle receives and executes it.Start the experiment from multiple positions relative to the target pallet.The experimental results show that only relying on the image signal of the monocular visible light camera can efficiently determine the driving direction in real time;when a person is approaching,it can stop and wait to ensure safety;the experimental vehicle can start from various positions relative to the target pallet to optimize The path completes the pallet picking process and has a high success rate.And it is low cost,flexible,easy to integrate,and has application value.