Surface Targets Recognition And Tracking Based On The Vision Of An Unmanned Surface Vessel

The unmanned surface vessel(USV)mainly depends on its vision system for surface targets recognition and obstacles detection.With the increasing intelligence of the USV’s vision system,it has a broad application prospect in military and civil fields,which makes the research on the surface targets recognition and tracking system of the USV becoming the current research hotspot.Under the complex sea environment and time-varying background,the tracking loss and inaccuracy of the targets to be identified and tracked are caused by the appearance and rotation changes.It is difficult to achieve accurate recognition and tracking only by traditional targets detection and tracking methods.Therefore,based on the current artificial intelligence technology,this thesis puts forward an intelligent solution for the recognition arid tracking of surface targets,which can identify and track surface targets in real time and accurately.First of all,in the detection of surface targets,there are some problems,such as the change of illumination conditions in surface scene and the difficulty of small target recognition in the distance,using the target detection algorithm of you only look once version 3(YOLOv3)to detect the surface targets.This paper analyzes the design principle of the feature map output network of YOLOv3 algorithm’s targets detection,and expounds the design mechanism of the non-maximum suppression filtering prediction box,the prediction of the target object boundary box and the loss function.The experimental results show that the recognition accuracy of the algorithm is higher and the recognition effect is better in the actual surface scene.Furthermore,in view of the shortcomings of minimizing the output sum of squared error(MOSSE)convolution filter tracking algorithm,which can’t track the surface target adaptively when the size of its shape and scale changes.An improved adaptive convolution filter tracking algorithm based on MOSSE is proposed,which updates the size of tracking box by adding scale change factor.The algorithm is robust to the changes of illumination,pose and occlusion in the tracking process of the surface target,and has the characteristics of real-time and robustness.The experimental results show that the improved algorithm is more accurate and reliable,which is very difficult in the actual tracking scene.Finally,in order to solve the problem that the first frame of image selected artificially in the MOSSE filtering method,the YOLOv3 targets detection algorithm is combined with the deep simple online and real time tracking(Deepsort)targets tracking algorithm to realize the autonomous recognition and tracking of the surface target.In this method,YOLOv3 is used to replace the detection link in the Deepsort,combining the motion and appearance information of the surface target and the Hungarian algorithm is used to solve the correlation between the predicted Kalman state and the new state.The Mahalanobis distance is used to evaluate it,and the cascade matching strategy is used to improve the matching accuracy,so as to realize the first detection and then tracking of the surface moving target.After the test in the actual surface scene,it has a good performance in the classification recognition and tracking effect of the USV.