Research On Vision-based Environment Awareness Technology For Unmanned Surface Vehicles

This thesis is based on the "huster-68" unmanned surface vehicles developed by Huazhong University of Science and Technology,focusing on the unmanned surface vehicles environment sensing technology.The position of the sky,the waterfront and the water surface can be obtained through the semantic segmentation of the water surface environment;the position and size of the object can be obtained through water surface object detection;and the water object can be classified in low-quality image which is caused by illumination or weather through low-contrast water surface object recognition.Then the above obtained information is transmitted to the central control system for autonomous cruise.The detailed research content of this paper is as follows:(1)In the semantic segmentation of complex water surface images,the traditional image segmentation method and the convolutional network segmentation effect are not elaborate enough.RGBP-FCN semantic segmentation network which combines polarization degree and full convolution network(FCN)is proposed,then the combination of polarization degree image and RGB image of the original scene is fed into RGBP-FCN to get the final segmentation result.Experiments show that the segmentation effect of this algorithm is more elaborate than traditional methods and full convolution networks.(2)In object detection of complex water surface,visual saliency algorithm combined with LAB color space and HSV color is proposed because the effect of traditional object detection method is easy to be influenced by water surface noise and convolutional neural network is unable to perform generic object detection.The object contour detected in LAB color space is relatively complete,but there is water surface noise interference.There is almost no water surface noise in HSV color space,but the object contour is incomplete.According to the advantages of the above two algorithms,the final detection result can be obtained by the decision level combination of the results of the above two algorithms.Experiments show that this algorithm can successfully detect objects of different scales,sizes and categories while suppressing surface noise.(3)In water surface object recognition of low-contrast image,the parameters of traditional image enhancement method multi-scale Retinex(MSR)need to be manually adjusted,so the convolutional neural network called REMSR is proposed.And the connection of REMSR and object recognition network is proposed,which can realize the end-to-end network of low-contrast images to object recognition label through the combination of low level task and high level task.In order to further improve the accuracy,this paper pre-trained the REMSR network and then fine-tuned it.Experiments show that the effect of object recognition of low-contrast water surface images has been significantly improved.