Marine Navigation Radar Target Detection with Multi-scale Spatiotemporal Feature Fusion-based Deep Learning

This article endeavors to tackle the challenge of radar target detection in dynamic scenarios characterized by strong maneuvering, such as those involving rotating radar carriers or high-speed moving targets. A deep learning model, termed the multi-scale spatiotemporal fusion network (MSSF-Net), is proposed to effectively tackle the aforementioned scenarios. The proposed method encompasses several modules, with a notable inclusion being the improved Faster R-CNN module, which incorporates attention and spatial transformation mechanisms. This integration enables the extraction of spatial offset features from the current frame’s echo image, facilitating the suppression of sea clutter and noise. Concurrently, the spatiotemporal feature fusion module within the network is designed to extract spatiotemporal information from historical frame echo images. It conducts multi-scale and multi-level feature fusion with the feature maps generated by other modules. This enables the capture of profound spatiotemporal features from a global perspective, thereby mitigating the decline in detection accuracy induced by the spatial shift of echo information resulting from strong maneuvering. Other modules within the method serve to manage the input and output of the model. Simulation results validate the enhanced accuracy of radar target detection achieved by the method under conditions of strong maneuvering.