Research On Object Detection And Tracking Based On Forward-looking Sonar

The exploration and development of marine environment depend more and more on the perception ability of the autonomous underwater vehicle to the surrounding environment.The forward-looking sonar(FLS)is just like a pair of eyes of underwater vehicles.The research of forward-looking sonar to detect and track underwater targets plays an important role in underwater environment exploration and operation.The FLS image is fundamentally different from the optical image in the model.It is of great significance to study the detection and tracking technology of the forward-looking sonar image.Based on the multi-beam forward-looking sonar,which can provide images with high frame rate and high resolution,this paper studies the target detection and multi-object tracking technology,and the main research work is as follows:In this paper,the detection and annotation method of FLS is carried out.An annotation method of object detection is proposed for the forward-looking sonar image specified,which can mark not only the object of interest,but also the object's shadow.An evaluation metrics of shadow detection accuracy is designed for this annotation method.Based on the improved method,a forward-looking sonar image generation experiment is designed to show that the annotation contains more target information.In order to enable the deep learning model to output both the object and shadow at the same time,the method of sonar object detection was studied.Firstly,the advantages and disadvantages of the two modification methods are analyzed.The two methods are easy to implement intuitively,but they cannot achieve high accuracy in both object detection and shadow detection tasks.In this paper,the model is improved based on these two methods,and the effectiveness of the improved method is verified by comparative experiments.For multi-object tracking problem of FLS,this paper has studied how to improve the tracking metrics when the carrier motion information is lacking.The rotation angle of sonar carrier is estimated based on the phase correlation algorithm.And then inputs the estimated value into the motion model as a control amount.The experiments on real sonar data provesthat the improved method is less sensitive to the algorithm parameters and has higher tracking performance.Finally,this paper combined with improved detection and tracking methods to carry out experiments,which verified that the improved detection model realized the tasks specified by the annotation method,output the bounding box of object and shadow at the same time,and also verified that the proposed tracking model improved the tracking performance.