| Unmanned surface vehicle is an important marine equipment and plays an important role in marine information exploration and offshore patrol.Context-aware technology is the key technology for the unmanned ship's normal driving,and the visual environment-aware technology is also an important key technology to determine whether the unmanned ship can be promoted in a large area.The most important task of the visual environment perception system is to detect the location of the obstacle target,and also includes tasks such as information recording and environmental detection.However,the sea surface visual environment sensing system is not mature due to environmental disturbances and other problems.The main problems are:when the obstacle target is small,because the sea surface is relatively empty,the proportion of the background area is too large,resulting in low target significance,which is prone to missed inspection;due to the influence of sea surface fog,fogging often occurs during imaging,resulting in Problems such as difficulty in detecting objects;target tracking problems in unmanned ships in complex environments.In order to solve these problems,the visual environment perception system usually contains many sub-modules.For this reason,the adaptability of each sub-module of the visual environment-aware system in the sea environment is deeply studied.This paper studies the visual environment perception system for the marine environment.According to the requirements of the project,combined with the characteristics of the sea surface environment,the overall scheme of the visual environment perception system is studied;according to the characteristics of the sea surface environment,the sea surface target region extraction technology is studied to obtain the potential target region,and the effect of reducing the invalid image region processing is achieved;Excavate the image information collected by the visual system,study the Haitian regional segmentation technology;to adapt the target detection algorithm to the characteristics of sea surface imaging,study the sea surface defogging technology,and study the anti-fog method of the target detection algorithm for the limitation of the defogging technology;At the same time,the unmanned ship environment sensing system can stably detect objects in motion and study the sea surface target tracking technology.According to the characteristics of the sea surface environment,the sea surface target area extraction technology is studied.The pictures collected by the unmanned ship in the sea during normal work have relatively empty features.Direct use of the original image for subsequent processing not only wastes computing resources,but also may cause the target area to be over-compressed and the target missed detection problem.To this end,firstly,the current mainstream target extraction method is adaptively studied,and the typical saliency detection algorithm is applied as a target area extraction method applied to the sea surface environment.In order to solve the problems of time-consuming and leaking alarms in the saliency detection method,the shallow convolutional neural network was used to test the effect of the method.In order to fully exploit the picture information collected by the visual system,the Haitian area segmentation technology is studied.Images acquired by unmanned vessels in the sea environment usually contain sky areas,while targets in the sky area are of no value to this topic.At the same time,the boundary area of the Haitian area is usually the area where new targets appear.Therefore,the study of the sea area segmentation technology has important value for the visual environment perception system.To this end,this paper studies the mainstream detection technology of Haitian boundary line based on Hough line detection,and deeply analyzes its existing problems.In order to solve the problem of weak detection ability of discontinuous sea boundary based on Hough line detection,a full-convolution network based on knowledge migration is proposed for sea-sky regional segmentation under the premise of meeting real-time requirements.In order to verify the effect,the test was conducted by experiment.Fog is the most important factor that interferes with the visual environment perception system.To this end,the anti-fog technology for sea surface target detection methods is studied.Experiment with the current mainstream defogging algorithm,analyze its dehazing effect and algorithm time-consuming.In order to solve the problems of high time-consuming defogging algorithm and weak defogging effect,the anti-fog method of target detection network is studied.In order to solve the problem of no anti-fog samples at present,the method of generating foggy samples based on the atmospheric single scattering inverse model is studied.In order to achieve the best anti-fog effect,the mainstream target detection model is firstly tested to test the effect of different visibility on its detection effect,and the appropriate energysample sequence is determined.The main target detection neural network is tested by different sample combination methods.In order to accomplish the target tracking tasks of various tracking scenarios of the subject,the target tracking technology is studied.Study the adaptability of the mainstream target tracking framework in the sea scene.Aiming at the characteristics of the anti-fog target detection network in this paper,the target tracking method combining target detection and Kalman filter is studied,and the experimental verification is carried out through the foggy scene video.Aiming at the problem that the target tracking method based on target detection network and Kalman filter can't deal with multi-target tracking scene,the target tracking method based on cyclic neural network is studied,and the tracking characteristics based on cyclic neural network are studied through various tracking scenarios.Finally,the target tracking framework for tracking the most appropriate tracking method for the current scene is designed by visibility. |
PDF Full Text Request