Research On Environment Perception Of USV Based On Stereo Vision And Filter Algorithm

With the development of water resources,Unmanned Surface Vehicle(USV)becomes the main surface transportation.The environment perception is the key to the next generation of USV,which enables USVs to gain the semantic understanding and state estimation of the targets.However,most of the current environment perception systems for USVs are not online for its high computational cost and not equipped with the ability of motion estimation for the target,which limits the autonomous and low-cost advantages of the USVs.This thesis studies the online perception system based on stereo vision and filter algorithm for small USVs.The system equips the USV with the semantic understanding ability and state estimation ability of the target in the environment.The main purpose of the environment perception system is to enable USV to have online environment analysis ability,which includes sparse semantic understanding ability based on stereo vision and state estimation ability based on hypothesis analysis filter.This thesis focuses on the low-cost perception system and completes the following work.(1)To extract semantic information from the environment at low cost: this thesis studies a sparse semantic extraction network that can work on GPU in real-time.It combines the YOLO and HED sub-networks to extract objects.And it proposes the Modified-ResNet and the semantic mask module for the semantic details.The whole network implements the sparse extraction of target semantic information at a low cost.(2)To solve the problem of scale ambiguity of visual information: this thesis introduces the scale information for the perception system through stereo vision,design a back-end optimization algorithm based on semantic information to reduce the mismatch,and complete the sparse semantic point cloud construction of the target object.(3)To solve the problem of target tracking: the thesis designs a radar-vision fusion module and the hypothesis analytic filter to make USV obtain the estimation ability of the target’s state for the target tracking.We propose a hypothesis analysis filter algorithm,which uses the‘nonlinear transfer’ noise model and estimates the moment of conditional probability density iteratively.In our experimental scenarios,the filter converges rapidly,is insensitive to the initial value and has good tracking performance.