Contingency Planning Of The Robot Vision Contamination Based On Deep-learning

With the rapid development of science and technology,autonomous mobile robots are gaining popularities in various fields which especially replacing humans to work in hostile and unknown environments.On one hand,to improve the ability adapting to complex environment,the robot needs to be equipped with the ability of environment perception like animals.On the other hand,due to the complexity and uncertainty of field environments,the robot needs to have the ability of contingency planning to deal with emergencies and improve reliability.For this,in this paper,self-adaptive oriented contingency planning of wheel-type mobile robots(WMRs)with chameleon-inspired perception was studied.A contingency planning was presented for the visual contamination of the mobile robot based on the framework of Case Based Reasoning(CBR)introspection in this paper,so as to enable the robot to have degraded functions.The hardware components of the robot mainly includes the upper computer vision system and the driving part of the lower computer.Based on the analysis of BioTRIZ the bionics principle according to the similarity between the chameleon and the robot vision system,a binocular PTZ vision system based on the imitation Chameleon was constructed and the coordinate system of the bionic vision system was established.For the stained robot cameras,the method of combining the frame-difference method with the Visual Background Extrector(ViBe)algorithm was proposed to extract the pollutants from the cameras of the mobile robot.And the morphological processing,filling,analysis of the connectivity,and the area and centroid of the contaminants were performed on the extracted contaminated images.The distance from the desired point to the boundary of the convex hull was calculated.The denoising process was implemented for the small area of pollution,and the entire imaging plane was divided into contaminant areas.The strategy design of the CBR solutions was made and the corresponding contingency plans under different pollution conditions were formulated.Based on the applied deep learning framework,the data sets were made and trained,and the training process of the network model was accomplished.Extract the pollution images and establish a CBR case database to analyze the three different situations of field-of-view stitching and develop a scanning strategy.According to whether the pollutants are blocked or not,the minimum cost of steering angle and focal length of the steering gear was calculated,and the camera motion optimization was performed.The algorithm was verified based on the 4WD4 WS mobile robot in IRAFS laboratory.Firstly,experiments were conducted using the pollutant extraction algorithm to verify the effectiveness of the algorithm in the different ground environments and lighting conditions.And then select the newly acquired image data to verify the effectiveness of the trained deep learning network.For different pollution situations while the transparency was greater than the threshold,the experiments of the CBR visual contaminated emergency planning were done in sandy and flat areas;and the effectiveness was verified.When the camera was contaminated and the transparency is smaller than the threshold,a camera cleaning experiment was implemented using the Coanda cleaning device;and the effectiveness of the cleaning device was verified.