Research On Surface Target Detection And Segmentation Algorithm Of USV Based On Deep Learning

Due to the complexity of the marine environment,unmanned surface vessel must have fast and accurate detection and recognition capabilities and adaptability to multiple scenarios when performing target detection and segmentation tasks.In the existing unmanned surface vessel surface target detection algorithm,the moving target detection algorithms such as background modeling are easily affected by the interference in the marine environment,which results in false detection and reduces the accuracy of target recognition.The target detection algorithm based on manual features is easy to generate window redundancy,and the calculation amount is large,which is difficult to apply to scenarios with strict real-time requirements.To solve the above problems,this paper proposes a target detection and segmentation algorithm based on deep learning,which uses the strong feature extraction and anti-jamming capabilities of depth convolution neural network to detect the sea target based on video images end to end,quickly identify and capture more image details,and provide accurate and realtime target information for subsequent unmanned ship collision avoidance operations or task conversion.The main work is as follows:1)A target detection algorithm based on the YOLO v3 network is proposed.The algorithm makes full use of the multi-scale features and the design idea of fast regression calculation to extract features of different resolutions from each position of the input image.The algorithm reduces the effects of disturbances such as waves,light and fog,and improves the accuracy and speed of identifying various target objects in the image.From the perspective of network versatility and scene complexity,a simpler and more practical ResNet-50 network structure is adopted to replace the original backbone network DarkNet-53,which is convenient for the implementation of subsequent joint algorithm.And based on the target objects commonly found in the marine environment,the training data set is made.2)When there are multiple target objects overlapping or too narrow in the image,the detection frame generated by the YOLO algorithm will appear offset or directly occupy the entire image,resulting in detection errors for the target position information in the image,thereby enabling Subsequent ship collision avoidance,path planning and other tasks produce misjudgments.Aiming at the above problems,a water surface segmentation scheme based on DeepLab v3 network is proposed in this paper,which uses four kinds of sampling rate dilated convolutions to resample the input feature map,and different scales of water surface pixel classification information are obtained from the captured image.The water surface in the image is segmented in real time to assist in judging the position contour of the target object and the basic navigational area.3)In order to solve the problem of low computational efficiency and difficult time synchronization caused by parallel algorithms in unmanned surface vessel,a joint algorithm of multiple tasks sharing the same feature extraction network is proposed,which can improve the efficiency of network parameter calculation and alleviate the unmanned surface vessel.And to ensure time synchronization of multiple algorithms.The joint algorithm adopts the detection and segmentation fusion data set,and improves the size of the target detection candidate frame according to the characteristics of the data set,reduces the calculation processing time,and improves the detection efficiency.The simulation results in the TensorFlow network framework verify the real-time and accuracy of the joint algorithm.Finally,the effectiveness of the joint algorithm is tested by a real ship test in a lake.