Multi-sensor Information Fusion And Perception Technology For Unmanned Surface Vehicles

As a highly intelligent carrier at sea,environmental perception is the core technology for unmanned surface vehicle to carry out most of its tasks and understand environmental information.When a single sensor depicts a target or scene,its perception ability is relatively limited,and it is limited to detecting information from a certain angle or focus.At the same time,the single unmanned surface vessel inevitably is ’ weak ’,and its perception range and efficiency are limited,which is prone to misjudgment and target loss.The results obtained are often one-sided,and the marine environment perception model cannot be fully constructed.Therefore,an algorithm framework of multi-modal information fusion perception technology suitable for multi-surface unmanned boats is designed for the marine environment,which enhances the fault tolerance and generalization of the perception system,and constructs a broader,perfect and accurate marine environment perception model.Firstly,the radar information on the unmanned surface vessel is processed,and the data characteristics of the lidar and the navigation radar are analyzed.For the marine environment,the conditional filter,the DBSCAN clustering algorithm and the Mean-Shift threshold algorithm are used to process the perception information,and the fusion is carried out on the basis of sensor calibration.An extended Kalman filter fusion scheme is proposed to fuse the data characteristics of the lidar and the navigation radar,and the CTRV model of the sea target motion is constructed and the fusion scheme is theoretically deduced.The simulation test verifies that the preliminary radar marine environment perception model is obtained.Secondly,focusing on the characteristics of multi-scale and diversity of sea surface targets,presented a dense-Yolov3-Tiny-SK target detection network to solve it.The network feature extraction part is improved to Dense architecture to strengthen the feature transmission of sea surface targets and effectively alleviate the problem of gradient disappearance.SK attention mechanism is introduced to strengthen the learning of neurons on the receptive field of sea surface targets at different scales.GIo U is used to optimize the loss function and improve the accuracy of border regression.The original network and the improved network are trained and tested by collecting real sea images.Finally,the experimental comparison verifies the superiority of the improved network.Then on the basis of the above multiple USVs information fusion and improve the marine environment perception model construction.Based on the calibration mapping between radar and visual sensor,the fused radar environment model and the detection information of optical image are fused at decision level,and two different fusion schemes and algorithm frameworks are proposed.Then the weighted model is constructed to evaluate the risk situation of the sea target with the attitude information and radar information of the ship.After the fusion of perception information between multi-boats,the fuzzy theory matching model is constructed,and two schemes and algorithm frameworks are proposed for different task requirements.The priority strategy is formulated,and the real target information is filtered and fused to construct a perfect marine environment perception model.Finally,the ’ QZ ’ series unmanned surface vehicles cluster is built in Zhanjiang sea area.The hardware foundation and sensor equipment of four unmanned boats are developed and used.Combined with the experimental data,the multi-sensor information fusion perception system is used to verify the field test.The obtained unmanned boat environmental perception model has good perception effect in the real marine environment,and realizes accurate perception of the marine environment.