Variable Scale Target Recognition Technology Based On Heterogeneous Sensors For Unmanned Surface Vehicle

As a high intelligent carrier on the sea,Unmanned Surface Vehicle(USV)is equipped with visual sensing and infrared sensing devices.Each sensing device has its own unique advantages,providing more abundant and comprehensive target information collection methods for the recognition of surface target,thus making a more accurate judgment on target information.However,the target scale changes caused by distance,observation angle and other factors in the course of navigation,lead to significant changes in the target features,which eventually affects the USV's "understanding" of the target information.In this context,this thesis aims to improve the intelligent level of USV itself by implementing research on recognition technology of heterogeneous sensors for variable scale target.During the registration for infrared and visible images,heat transfer will cause the blurring at the black-and-white boundary of checkerboard.To solve this problem,we designed a simple and low-cost circular calibration device for infrared and visible dual cameras based on the property of uniform heat dissipation around the spherical heating object.The calibration test result shows that the back projection error is about 0.15 pixels,which verifies the rationality and practicability of the calibration platform designed.In this thesis,we focus on two difficulties of feature fusion of heterogeneous sensor,which are feature consistency and optimization of fusion strategy.On this basis,we built a dense connection network to extract the feature information of the target,weaken the physical meaning difference between the visible image and the infrared image,reduce the deviation of the fusion image in the spatial position,and finally realize the feature-level fusion.Besides,we proposed two new fusion strategies,weighted average fusion and improved serial fusion,which have achieved a better fusion effect based on the property that the sparse nature of L1-norm can make a reasonable selection for feature information according to its own characteristics during the fusion.The conventional neural network uses a single sensor to solve the variable scale problem of the target,which has the defect of unitary information about the target features.But the feature pyramid network designed based on deep fusion feature of heterogeneous sensors,can enhance the robustness of networks to the changes of target scale.Taking the advantage of infrared sensor that can detect small targets far away,the fusion network of low-level features using new weighted average fusion strategy is improved,which can get a higher overall target recognition accuracy.The result of variable scale target recognition shows that the effect of heterogeneous sensors is better than that of any single sensor,and realize the advantage complementation of each sensor.In the actual marine environment,we took the "Robot-X" as the test platform to carry out field experiment on the target recognition network which adopts the new weighted average fusion of low-level features.The experimental results show that the target recognition network can improve the overall accuracy and real-time performance of the whole feature fusion system,reduce the false recognition and mismatching caused by the interference of environment and non-target factors,improve the accuracy of target recognition under the influence of complex marine environment,significantly enhance the intelligent comprehensive perception ability of USV,and realize the true meaning of "visual sense".