Research On Target Detection And Region Segmentation Of Amphibious Unmanned Vehicle Based On Vision

China is busy shipping,water accidents and illegal berthing of vessels,and a series of problems are becoming more and more serious.Compared with the ship,the water-air amphibious aircraft has better vision and can reach the monitoring area faster than the ship.Therefore,it is of great significance to complete channel monitoring and vessel management based on image processing on amphibious unmanned aerial vehicles.The navigation channel monitoring and vessel detection based on computer vision are needed for amphibious unmanned aerial vehicles,but the existing detection algorithm model based on depth learning is too large to operate the embedded equipment well.In order to solve the above problems,this thesis proposes a water surface target detection algorithm and a water surface region segmentation algorithm based on model compression,and designs an amphibious vehicle water surface monitoring management system to realize the task of channel monitoringFirstly,in order to solve the problem that the classification accuracy and the model are too large in the image target detection task,three improvements are proposed for yolov3: 1)the depth separable convolution and lightweight attention structure are introduced,and the extracted features are fused with multi-scale features,so as to reduce the size of the model,reduce the parameters and meet the needs of real-time detection.2)Kmeans + + is used to set the anchor box of data set in advance to improve the detection accuracy.3)Using mish activation function instead of leakyrelu can avoid gradient saturation due to capping and improve the overall accuracy.The data set is enhanced to better adapt to the detection task of water surface targets under different weather conditions.Finally,the experimental platform is built to compare the improved yolov3 model with the original yolov3 model,and verify the real-time and accuracy of the algorithm.Secondly,in order to ensure the accurate and fast region segmentation of the images taken by amphibious vehicles,the model based on deeplabv3 + is improved to meet the requirements.The specific improvements are as follows: 1)referring to the inverse residual structure in mobilenet and using the average pooling to replace the full connection operation,the main feature extraction network is designed to realize the feature extraction of shore,ship and water Accurate segmentation of the junction.2)By replacing the standard activation function in the network,the segmentation accuracy is improved.3)The size of the input network is limited to accelerate the model training and prediction.Compare the improved deeplav3 + model,xception based deeplav3 + model and segnet model to verify the size and segmentation accuracy of the improved deeplav3 model.Finally,the amphibious vehicle surface monitoring and management system is designed to achieve the following functions: 1)planning the monitoring area of the vehicle;2)display the status of the vehicle;3)identification,detection and display of the surface vessels;4)segmentation and display of the water surface,vessels and land.