Shuyue Li, Junjie Wang, Jinlu Sheng, Ziyu Liu, Shixin Li, and Ying Cui
Unmanned surface vehicle, target detection, deep learning, lightweight network
Maritime target detection is a critical component of navigation safety for unmanned surface vehicles (USVs), particularly under foggy weather conditions. We propose an efficient and lightweight method for maritime target detection suitable for foggy weather conditions. This approach aims to address the high cost of data acquisition and to enhance the target detection effectiveness. The proposed method involves several steps to enhance the detection effectiveness and efficiency. Firstly, we improved the accuracy of foggy image synthesis by formulating a more realistic loss function for the GAN model. Secondly, we reduced the model size and number of parameters by introducing depthwise-separable convolution instead of conventional convolution. Finally, we applied a lightweight backbone to improve the high-dimensional maritime target features and accelerate the inference speed. The experimental results demonstrate significant improvement in the accuracy and efficiency of the proposed model. Our model achieved an average accuracy of 89%, which is a significant improvement over the 77% accuracy of the original YOLOv4 model. Additionally, the computational volume of our model was reduced by 83%, and the real-time detection speed reached 45.8 frames per second. This improved accuracy and efficiency make the proposed method more appropriate for complex conditions and enhance the safety of USV navigation.
Important Links:
Go Back
