Real-time Target Detection And Recognition Of Unmanned Underwater Vehicle Based On Vision

The deep sea area accounts for 90% of the global ocean area,and more than half of international trade is transported through the ocean.While the ocean promotes human civilization and economic development,it also leaves behind numerous human sites and relics.China has a glor IoUs history of maritime trade in ancient times.On the one hand,deep-sea relic detection and excavation can promote the development and progress of advanced marine equipment.On the another hand,it is also significant for inheriting and carrying forward traditional culture.In order to meet the needs of underwater archaeological robots for near-seabed careful target detection,this dissertation proposes a vision-based real-time underwater target detection and recognition.Based on the key frame extraction and multi-granularity model compression algorithm,this dissertation aims to realize the rapid and accurate target detection of underwater archaeological robots,and develops the research that real-time target detection and recognition of underwater archaeological robots based on vision.First of all,this dissertation summarizes the existing underwater archaeological patterns,equipments and the application status of unmanned underwater vehicles in underwater archaeology at home and abroad,focuses on the underwater target detection methods of unmanned underwater vehicles,compares the difference between underwater acoustic target detection and optical target detection,and points out the importance of vision-based underwater target detection and recognition in deep-sea archaeology.Secondly,this dissertation introduces the principles and implementation steps of traditional target detection and deep learning-based target detection algorithms.Comparative analysis points out the inevitability of using deep learning target detection algorithms,and introduces var IoUs evaluation indicators in the deep learning detection model.According to the backbone network architecture,two types of deep learning target detection algorithms(one-stage and two-stage)are analysed,and the model of the R-CNN series,YOLO series and Anchor-free series algorithms are selectively analyzed.At the same time,the common data sets for target detection are used to var IoUs target detection algorithms to analyze the advantages,disadvantages and applicable scenarios.And making a theoretical foundation for the selection of basic models in the following.Thirdly,according to the characteristics of underwater relics,the target detection objects of underwater relics are determined.In order to obtain an accurate detection model,an underwater relics data set is established named Underwater?relic,and the original data set is expanded by unsupervised and supervised data enhancement methods.Moreover,to solve the problem of image distortion in the deep sea that collection of unmanned underwater vehicles,a real deep-sea image data set is established,and deep-sea image imaging models are studied.Based on the characteristics of deep-sea image imaging,a linear scene depth model is proposed.Transmission map and background light of the original image can be achieved based on scene depth model,combining with the underwater imaging model,the original image can be restored rapidly.Then,the actual requirements for real-time target detection and recognition of underwater relics of unmanned underwater vehicles are analyzed,and a fusion-based key frame extraction algorithm and a multi-granularity model compression algorithm are proposed.Using structural similarity,color histogram and image entropy to extract the key frames of the video collected by the unmanned underwater vehicles.At the same time,under the premise of ensuring accuracy and structural integrity,the multi-granularity model compression algorithm of nuclear pruning and inter-layer pruning is used to perform bidirectional model compression on YOLOV4 to realize real-time target detection and recognition of underwater underwater vehicles.Finally,in order to meet the requirements of the space and power consumption of the unmanned underwater vehicles,the Nvidia Jeston TX2 image processor was selected as the underwater relic target detection platform.Transplant the detection model after using high-load,high-power equipment training and parameter adjustment to the TX2 embedded platform for real-time target detection results and speed testing of underwater relics.Experiments show that the detection speed of YOLOV4 after compression is 6.27 times that of uncompressed,and by extracting key frames from the input video of the compressed model,the detection speed of the model can be increased to 18.65 times of the original.In summary,the target of this dissertation is the real-time target detection of underwater archaeological robots.This dissertation establishes the underwater relics dataset Underwater?relic and enhances its data,completes the design and experiment of the online fast deep-sea image restoration algorithm,creates the key frame extraction algorithm of fusion ideas and the model compression algorithm based on multi-granularity pruning are realized,and realizes real-time underwater relic target detection and recognition on the embedded platform.This dissertation provides design scheme and technical support for underwater robot to carry out deep sea accurate and autonomous archaeology based vision.