Research On Vision Target Detection Method Of Robot’s Electrified Lap Drainage Line

Live working in distribution network is an important operation mode to improve power supply reliability,reduce power outage loss and ensure power grid safety.At present,domestic live working is still dominated by manual work,which is easy to cause personal casualties.It is an inevitable trend to develop live working robots to replace it.The "what you see is what you get" robot vision guidance system can ensure that the mechanical arm can complete the live lap drainage wire work safely and efficiently.Therefore,the problem of accurate detection of the operation target vision in the live overlapping drain wire operation needs to be solved urgently.The difficulty of vision detection in the process of electrified lap drain wire is the accurate detection of the operation target position in the process of stripping and wiring.In the process of stripping,it is necessary to detect the predetermined stripping position,and in the process of wiring,it is necessary to detect the exposed area of stripped overhead cable.(1)Stripping position detection.Deep learning network has strong ability of feature extraction,which is the most commonly used target detection method at present.However,it is difficult to give full play to its advantages because there is no obvious feature and uncertainty of the location of the stripping position,so it cannot be directly detected.To solve this problem,this paper uses Faster R-CNN(Faster Region-based Convolutional Neural Networks)and MTCNN(Multi-task Cascaded Convolutional Networks)to detect the indirect target,and then combines the spatial position conversion to detect the predetermined stripping position.(2)Detection of exposed areas.There is regular texture in the exposed area,which can be detected by using this feature,but the traditional texture filtering needs prior texture parameters,which are difficult to obtain prior texture parameters due to the complex background and uncertain target attitude in the wiring process.To solve this problem,spectral clustering and Gabor filter are used to detect different attitudes in the complex background without prior texture parameters of exposed areas.The main work and contributions of this paper are as follows:(1)The vision module and the main wire frame of the vision guidance experiment platform of the live working robot are designed and implemented.In order to collect the experimental data,verify the feasibility and effect of the method proposed in this paper,according to the operational standards,using the modular design method,the visual module and the main line frame are designed and implemented.(2)Based on the improved depth learning network and the spatial coordinate conversion,the detection method of striped line position is proposed.Aiming at the problem of insufficient feature extraction ability of P-Net(Proposal Network)and R-Net(Refinement Network)in MTCNN,the Faster R-CNN is used to replace them,and aiming at the problem that the RPN network in Faster R-CNN does not extract the candidate frame accurately,K-means algorithm is used to cluster the candidate frame,and the network detection accuracy is improved by 4.3%.Aiming at the problem that the O-Net(Output Network)detection effect of MTCNN is unsatisfactory,this paper improves the detection accuracy of the network from two aspects of training optimizer and model fusion.Combined with the spatial position conversion,the final detection gets the predetermined stripping position.The L1 error of stripping position is 8.96 mm,L2 error is 6.26 mm,which is less than the 2cm positioning error required by the project.(3)The detection method of exposed area of overhead cable based on Gabor filter is proposed.To solve the problem that the traditional texture detection needs prior texture information,this paper proposes a detection method of bare area of overhead cable using spectral clustering and Gabor filter,which can accurately extract the location of exposed area under different background and different posture,and the positioning accuracy is higher than that of the contrast Mask R-CNN(Mask Region-based Convolutional Neural Networks).In summary,this paper designs a method of vision target detection in the process of robot electrified lap drainage wire work.Through the self-made robot vision guidance experiment,the two algorithms are tested and verified.The experimental results show that the two algorithms can effectively detect the work target,verify the validity of the target detection method of the robot live lap drain wire work,and support the project related topics to enter the pre acceptance stage.