Research On Object Detection Of USV Based On Images

With the development of artificial intelligence technology,the research and layout of unmanned system,unmanned surface vehicle(USV)as a major branch of unmanned system,which is the development trend of future shipping.USV achieves full intelligence that can replace people with performing dangerous or long-term missions to reduce personnel casualties and costs.Object detection is a key technology for USV to achieve full intelligence.In recent years,the object detection algorithms based on the convolutional neural network(CNN)have emerged as the advanced technology in its research field.The object detection algorithms based on deep learning are applied to USV,which makes the environment-sensing task of USV to be completed more efficiently and intelligently.This paper aims at image-based object detection of USV.By analyzing the impacts of marine environment and characteristics of objects,the general plan for object detection of USV is designed.The water boundary detection and object detection methods are discussed in detail.The main contents are as follows:Firstly,through the analysis of marine environment and characteristics of objects,the overall scheme of the water boundary detection method and object detection method based on image processing are designed.According to the real-time and accuracy requirements of the USV object detection system,the image acquisition hardware device selection and software environment selection and configuration are completed.Secondly,the water boundary detection method is studied.By analyzing the features of water surface images,the highlight region removal and median filter are selected as preprocessing for the light interference and noise problems.Then analyzing the existed water boundaries detection methods,the detection method based on gray-level co-occurrence matrix texture entropy and morphological open operations is proposed for the problem of poorly-detected coastlines with complex backgrounds.Experimental results show that the proposed method based on texture entropy has higher detection accuracy than other methods and can be applied to the detection of sea-sky line and coastline at the same time.Thirdly,The object detection methods based on convolutional neural networks are analyzed and compared.Marine images data is prepared,for the problem of less data samples,and the theory and application of migration learning are discussed,and the structure principles of main network VGG16 and ResNet are introduced.Three detection algorithms Faster R-CNN,YOLO and SSD are analyzed in detail,The experiment compares the detection effectiveness of the three algorithms on the sea images.The result shows that the Faster R-CNN algorithm based on the ResNet is more suitable for the object detection of USV.Finally,aiming at the problem that Faster R-CNN has a low detection accuracy on tiny objects,the detection method of ResNet and DenseNet hybrid master network combined with cyclic feature pyramid is proposed.Low-level multiple detail features and high-level strong semantic features were recurrently used to reinforce the detection for tiny targets.The experimental results show that the proposed method has higher means average precision than the other detection methods for tiny objects,and is more suitable for USV object detection applications.