Moving Target Detection in Sea Clutter Based on Deep Learning Methods

Abstract

Moving target detection in sea clutter is always a challenging problem due to the influence of sea clutter. Traditional methods mainly solve this problem by designing special feature- based detectors. By using specially designed features such as fractal features and time-frequency features, the detection question can be converted to a classification problem. Since the detection performance is directly related to the extracted target features, how to choose proper features should be considered carefully, which requires much expert knowledge. Recently, deep learning methods have achieved great success in the task of extracting object features and detecting desired objects in computer vision field. In this paper, we proposed a novel marine target detection method based on Yolov4 network and Kalman filter to extract target features and detect targets automatically. We built our own dataset based on radar data measured at Yantai Harbor. By analyzing the characteristics of our own dataset, we combined Yolov4 network with Kalman filter to improve the detection performance. Besides, we carried out further experiments to compare the performance between YOLOv3, YOLOv4 and YOLOv4-Kalman methods. The experimental results on real-scene dataset prove that our proposed method shows better detection performance in accuracy compared with other detection methods.