Maritime Target Recognition Based On Multi Camera

With the development of science and technology in recent years,deep learning technology has gradually entered the public’s field of vision.Aiming at the challenges brought by the use of multi-sensor fusion information tasks in intelligent ship navigation situation awareness,a multi-camera-based maritime target recognition method has been proposed.Traditional marine target recognition basically adopts the way of human eye observation.The results obtained by this method are greatly influenced by the subjective factors of the observer,and the observation range and observation distance are small at the same time.The multi-camera-based maritime target recognition method introduces a deep learning model in target recognition,which makes the recognition result more objective,accurate and scientific.This technology uses multiple cameras to obtain images of a wide range of scenes around the ship at the same time,improving the efficiency of observing targets.This paper first designed a system composed of multiple cameras with the same model and fixed position and angle,and studied the image preprocessing methods suitable for this situation based on the actual images taken by the camera on the sea.Then analyze the characteristics of the marine target,divide the marine target into 7 small categories,and use web crawlers,real ship shooting and data augmentation to obtain specific marine target data set,And the data set is processed into the style of the database required for standard network training.Then study the image stitching algorithm based on multiple cameras,According to the different methods used in each step of image mosaic,the specific theory and actual experimental results are studied and compared,and the most efficient combination method is selected.Then analyze the detection accuracy and detection rate of various target recognition networks,select the YOLOv3 network model with the best performance compatibility for the above two aspects at this stage,and use the self-made marine target data set for training.The average target detection accuracy(m AP)of all categories finally obtained reaches 91.85%,which meets the requirements of experimental target recognition and detection accuracy.And compare and analyze the influence of mirror augmentation method on the accuracy of model target detection.Finally,analyze the impact of different stitching algorithms on the accuracy of target recognition.Starting from the overall scheme of maritime target recognition based on multiple cameras,this paper studies the classification and collection methods of maritime targets,image stitching algorithms,and maritime target recognition algorithms,and completes the realization and verification of the overall algorithm.Experiments show that for multi-camera systems to obtain multiple contiguous images,Using the image stitching and YOLOv3network-based maritime target recognition method proposed in this paper,the target category and position information in the stitched image can be effectively obtained.The experimental results meet the preset index requirements.