Research And Implementation Of Visual Detection Technology For Pipeline Inspection Robot Based On SS-YOLO Algorithm

With the fast development of urbanization,urban underground pipe gallery has been well developed.The defects of urban underground pipe gallery are closely related to the safety of people's lives and property,so the defect detection of urban underground pipe gallery is very important.Traditional detection requires manual observation of the real-time video returned by the robot,the cost of detection is high and the accuracy of detection cannot be guaranteed.The object detection algorithm can detect the defects in the pipe gallery in real time without manual observation,and the detection accuracy is high.This thesis,by improving the feature extraction network and the method of obtaining a prior box,a new object detection algorithm is proposed for the detection of urban underground pipe gallery.The computing power of the pipeline inspection robot is lower than the computer's graphics processing unit.In order to solve the problem that the object detection algorithm cannot guarantee the detection accuracy when the detection speed meets the requirements,this thesis proposes a new object detection algorithm for the detection of urban underground pipe gallery after the improvement of YOLOv3-Tiny.The main research contents are as follows:In order to solve the problem that the low detection accuracy of YOLOv3-Tiny,this thesis improves its feature extraction network.Based on the Shuffle Net V2 network structure,the weight of each channel is obtained according to the importance of each channel in the feature map,the feature map is recalibrated in the channel dimension by multiplying the weight with each channel in the feature map.Through the above improvements,the Shuffle?Senet module is obtained.Based on this module,the feature extraction network was created.In this thesis,the kmeans++ algorithm is used to obtain the prior box,and the distance will be considered when initializing the clustering center to avoid the problem of local optimization.Through the above improvements to YOLOv3-Tiny,a new algorithm SS-YOLO is proposed to detect defects in the urban underground pipe gallery.In order to solve the problem that few data set can't meet the needs of training the object detection algorithm,this thesis uses deep convolutional generative adversarial networks model to generate the defect samples to increase the data set.This thesis extracts part of data from the data set to train the deep convolutional generative adversarial networks model,and after training generates data set with the same characteristics as the original data set,enriching the training set of the object detection algorithm to improve the accuracy of detection.According to the experimental results,the improved SS-YOLO's frames per second(FPS)is 4 frames higher than YOLOv3-Tiny,and the mean average precision(MAP)is 6.5% higher,the speed and accuracy of the detection are both higher than the previous algorithm.In the detection of urban underground pipe gallery,the mean average precision of SS-YOLO algorithm detection in this thesis has reached 0.895,which is improved compared to 0.735 of YOLOv3-Tiny.It can be seen from the experimental results that the SS-YOLO proposed in this thesis can meet the needs of defect detection of urban underground pipe gallery.