基于DFI一体化网络的水下抗干扰目标跟踪方法

Due to the uniqueness of the underwater environment, the AUV tracking control effect is easily affected by the complicated underwater environment, such as water flow, visible light reflection, etc., which interferes with the performance of the target detector and reduces the effect of tracking control algorithm. To solve the problem, the detection and feature extraction integration network (DFI Network) is proposed. Based on the traditional YOLOv3 detection network, a feature extraction network is designed to output the target feature information by convolution and maximum pooling operations, and the state vector dimension is expanded to 10 dimensions. At the same time, augmented Kalman filter is constructed for the extended dimensional state vector, and the target tracking of AUV in the interference environment is realized by subsequent matching and tracking controlling. The proposed algorithm is tested using some annotated underwater target tracking datasets and compared with the original YOLOv3 algorithm. The results show that the proposed algorithm improves tracking accuracy by more than 30% with a smaller frame rate loss of 2FPS compared with the original YOLOv3 algorithm.