Research On Target Detection Technology Of Multi-classifier Fusion Of Marine Radar Based On HOG Feature

With the proposal of the strategy of maritime power,the development of maritime ship navigation is particularly important for the protection of our country’s maritime territory.X-band navigation radar is widely used in target detection in the field of ship navigation due to its high resolution,high reliability,and small blind spot for distance detection.In maritime target detection,traditional target detection algorithms mostly use constant false alarm rate detection technology to detect radar images point by point and line by line,which requires a lot of computing time,cannot meet the requirements of real-time detection,and is limited by the high resolution of radar,the detection efficiency is low.Aiming at the problems of high time complexity and low detection efficiency in the detection process of traditional target detection algorithms,this paper constructs a detection process combining coarse detection and fine detection,and improves a fast target detection method based on HOG features.The measured data of the scientific research project"Shipborne X-band Navigational Radar Inversion Technology of Waves" has been verified.Before detection,it is necessary to preprocess the radar image and select appropriate feature to characterize the difference between the target and the non-target in the radar images.In this paper,from the perspective of image vision,each radar image is divided into block sub-images,and the X-band marine radar sample features library are established,and according to the different gradient information between target and non-target,a HOG feature based on gradient histogram is proposed.It is verified by experiments that this feature can distinguish targets from non-targets well.At the same time,for the feature dimension is too large,feature engineering is used to process the feature to prepare for the subsequent detection process.In the pre-detection process,we use three classification algorithms to classify images,namely the Support Vector Machines(SVM)algorithm,the Random Forest(RF)algorithm and the K-Nearest Neighbor(KNN)algorithm.Then,the target detection of each classification algorithm is studied.In the process of pre-detection with each classification algorithm,the corresponding algorithm is used to train the constructed image sample feature library to obtain a classification model,and then the radar image to be detected is segmented and sent to the classification model for classification.The sub-images containing the target are filtered out to reduce the detection time.Finally,the detection performance of each detector is compared and analyzed,and the result with the best detection performance of the detector is sent to the subsequent fine detection process.In the fine detection process,the block sub-images containing the target in the pre-detection result are detected.Then the measured data is analyzed to obtain the detector threshold,and the CFAR detector is used for detection to obtain the specific position of the target.All detected sub-images are stitched together to form the specific orientation of the target and non-target of the entire radar image.At the same time,each complete radar image is detected point by point and line by line by using the classical detection algorithm.Finally,in view of the different classification and detection performance of each classifier,a method based on Stacking fusion is proposed to fuse the three classifiers into an optimal classifier[58].Experiments show that the results obtained by the pre-detection and fine detection method adopted in this paper are shorter than the classical detection methods,and the detection accuracy is higher.In the meanwhile,the stacking fusion method can further improve the classification detection accuracy in the pre-detection process.