Research On Target Pointing Control Technology Of Arm-type Launch System Based On Machine Vision

Accurate perception of water level information changes is one of the key links to achieve fine water control and flooding,while existing technologies cannot meet the needs of water level recognition in complex harsh scenes such as night,haze,rain,snow,floating object occlusion and shadow.In view of this,this paper systematically carries out research work on intelligent water level detection for scenes with and without water rulers based on in-depth analysis of the characteristics of water level images in complex and harsh environments,making full use of the ability of deep learning networks to characterize semantic features:(1)In the scene with water ruler,the transformation of image coordinates to world coordinates can be completed by making full use of the water ruler scale,and the detection for the water ruler scale is typical of small target detection.In view of this,a water level detection method that integrates improved YOLOv5 and Rank SE with water scale is proposed: first,a multi-level feature fusion method that strengthens small-scale features is proposed to improve the YOLOv5 algorithm to enhance the ability to capture small targets;then,the Rank SE module is incorporated to further enhance the perception of small targets;finally,a new water level height solution method is proposed that only Finally,a new water level height solution method is proposed,which can obtain accurate water level elevation information by using only part of the water gauge anchor frame information,which greatly improves the robustness of the detection method.The relevant experimental results show that the water level detection accuracy of this method reaches 98.5% under the complex and harsh scenarios with water gauge.(2)For water level detection in scenarios without water gauge,exploratory research is carried out based on two technical approaches,target detection and semantic segmentation,respectively.In the study of water level without water gauge based on target detection technology,an intelligent water level detection method integrating improved YOLOv5 and Kalman filter principle is proposed.The core technologies include:(i)introducing YOLOv5 to detect the water level line(water shore demarcation line)and using linear fitting method to obtain the actual water level line,while the unconventional method is based on the water surface for identification;(ii)targeting the water level line in the extension direction(ii)for the water level line in the extension direction is infinitely large and in its normal direction is infinitely small,a multi-level feature fusion method to strengthen the mesoscale features is proposed to improve the original YOLOv5 algorithm;(iii)the Kalman filter is used to introduce the water level history information as a priori knowledge to improve the generalization performance of this technology to complex and harsh environments;(iv)a fixed marker in the image is added to the deep learning network for training,and the actual water level is solved according to the real size of the marker elevation to achieve a water-rulefree detection scheme.Relevant experiments and practice show that the improved YOLOv5 is more lightweight;the slope accuracy of this water level intelligent detection technology without water ruler is 97.3%,which is 2.4% higher than the original YOLOv5 algorithm;the intercept accuracy is 99.3%,which is 0.5% higher than the original algorithm;in night,haze,rain,snow,floating object occlusion,shadow and other complex and harsh environments can automatically and accurately identify the water level elevation with the error less than 0.1m.(3)In the study of water level without water scale based on semantic segmentation technology,a Unet model(TRCAM-Unet)is built to fuse Transformer and residual channel attention mechanism,and then the intelligent detection method of water level without water scale based on TRCAM-Unet for harsh scenes is proposed,and the core technologies include multi-level feature fusion through the full-scale connection structure,and the The core techniques include multi-level feature fusion through the full-scale connection structure,strengthening the correlation of regional features through the Transformer module,and strengthening the representation of useful information and weakening the interference of useless information through the residual channel attention module.Related experiments and practice show that in the water level detection without water scale in complex and harsh scenes,TRCAM-Unet achieves 98.40% MIOU score and 99.20% MPA score,and the maximum error of water level detection at a distance of about 150 m does not exceed 0.08 m,and the mean value of water level deviation(MLD)is only 1.609×10-2m,which is better than Deeplab,PSPNet and other conventional semantic segmentation algorithms.All the above three methods can effectively adapt to complex harsh scenes such as night,haze,rain,snow,floating object occlusion and shadow.Among them,the water level detection method fusing improved YOLOv5 and Rank SE is applicable to the scenario with water level detection;the water level intelligent detection method fusing improved YOLOv5 and Kalman filter principle is applicable to the scenario without water level detection where the water level line is gentle;and the water level intelligent detection method based on TRCAM-Unet is applicable to all scenarios without water level detection.The research results have important application value for solving the problem of accurate water level detection under severe scenarios and flood disaster warning.