Research On Algorithm Based On Monocular Vision For Surface Targets Detection And Location

Unmanned surface vehicle needs to sense the surrounding environment in real time and avoid obstacles autonomously by acquiring obstacle information within the field of vision.Visual sensor has advantages such as good real-time performance,simple layout and low cost.In this paper,aiming at the practical work of the unmanned surface vehicle,a method of surface target localization based on monocular vision is proposed to improve the detection accuracy of multiple targets on the surface.This method requires the target detection algorithm to obtain the specific position information of the target in the image.As a kind of mature bipolar detection algorithm,Faster R-CNN has better precision and portability of border regression.Based on this,the concrete work done in this paper is as follows.Firstly,the sailing environment of USV is evaluated and a multi-category surface target data set is constructed.By comparing the detection effects of three classical convolutional neural networks(VGG16,Google Net and Res Net50)on the surface target dataset,Res Net50 was selected as the backbone feature extraction network.The analysis of detection results of different convolutional neural network detection models shows that the original Faster R-CNN algorithm misses more detection of small-scale targets in the dataset,resulting in poor overall detection accuracy.To solve this problem,the original algorithm is improved in this paper.Firstly,the feature pyramid network integrated with Squeeze-and-Excitation mechanism is used to replace the top-down feature output method of Faster R-CNN algorithm.Secondly,the Online-Hard-Example-Mining strategy is used to replace the positive and negative sample training methods in the original Faster R-CNN algorithm.The experimental results show that the improved fusion algorithm can effectively improve the overall detection performance of the surface target dataset.Then,the application effects of the original target detection algorithm and the fusion improved algorithm in the monocular vision positioning method were compared respectively.The positioning experiment was carried out in the multi-target navigation video on the surface shot by the fixed camera.The results show that the optimization of the target detection algorithm can effectively improve the positioning accuracy of monocular vision,but the effect of this positioning method is slightly poor for long-distance targets,and the overall positioning effect is well.Finally,based on the navigation video of the cooperative working vessel collected by the "5800 unmanned surface vehicle",the practicability of the monocular vision positioning method proposed in this paper in the real working environment of unmanned surface vessel is explored.The test results show that the proposed monocular vision method can better finish on the surface of the real environment target detection and localization task.It is of positive significance to assist unmanned surface vehicle in obstacle avoidance and real-time path planning.