| Currently,the application and research of mobile robots combined with mechanical arms in uncertain environments is not yet mature.Due to the involvement of multiple disciplines such as visual detection,real-time control,and mechanical arm kinematics,research on machine vision-based mechanical arm hand-eye coordination remains one of the difficulties in the field of mobile robots.The existing detection algorithms are mainly based on deep learning methods,which bring about problems such as complex network structure,large model parameter quantities,and poor detection efficiency.This paper studies the object detection algorithm based on YOLOv5s as the benchmark model,optimizes the network structure,parameter quantity,and computational complexity,improves hand-eye calibration and image enhancement,and has important theoretical and applied value for improving the efficiency of mobile robotic arm hand-eye coordination.The specific contents are as follows.To address the problems of complex network structure,poor timeliness in door handle detection,and large model parameter quantities in the YOLOv5 algorithm,improvements are proposed for the backbone and Neck networks.Firstly,ShuffleNetv2 lightweight network is used to replace the backbone network,and Hard-swish activation function is used to replace ReLU,which improves detection accuracy and speeds up computation.Secondly,dilated convolution kernels are used in the network modules,and the network is pruned to reduce parameter quantities and computational complexity.Additionally,CBAM attention mechanism is added to the Neck network,global pooling is replaced with an adaptive method,and the LeakyReLU activation function is used to replace the activation function in the shared fully connected layer,thus reducing the gradient backpropagation stagnation phenomenon and accelerating network training.Experimental results show that the improved detection algorithm has improved precision,recall,mAP0.5,and mAP0.5:0.95 performance by 6.4%,4.1%,5.1%,and 5.3%,respectively.Moreover,its parameter quantity and computational complexity are reduced by 50%compared to YOLOv5s,the training time is reduced by 50%,and the inference time on mobile devices is reduced by 5-6ms.Due to the problem that it is difficult to perform joint calibration by fixing the calibration board at the end of the mechanical arm,a fixed viewpoint hand-eye calibration method based on two-dimensional codes is proposed.Combining image processing and two-dimensional code recognition and positioning technology,the edge detection operator is used to process the acquired image data to obtain the boundary lines of the two-dimensional code positioning area.Then,the corner positions of the two-dimensional code are obtained through coordinate transformation and EPnP for joint calibration.The improved calibration method reduces the average error by 16%compared to traditional calibration methods.To address the problem of dark or blurry images caused by lighting conditions or incomplete camera aperture closure during image acquisition,improvements are made to the Zero-DCE-Net image enhancement algorithm.The introduction of deconvolution and residual block networks compensates for the loss of image detail information during convolution and retains image details while achieving good denoising results.This improves the generalization of the network model and helps improve the target detection accuracy of hand-eye collaboration.Using image evaluation methods PSNR and SSIM as metrics,the image quality is improved by 18%and 5%,respectively.Finally,testing on an epidemic prevention and disinfection robot shows that the accuracy,inference time,and frame rate are 99%,25ms,and 40FPS,respectively,meeting the requirements of practical applications. |
