Multi-Modal Structured Pruning for USV-Based Waterway Detection Based on Radar-Vision Fusion

Radar-vision fusion, with more reliable performance at lower cost, has been widely used in autonomous vehicles. In the waterways, the perception of unmanned surface vessels is essential for autonomous navigation. However, the large amount of computation increases the demand for high-performance computing devices and causes severe power consumption. To reduce the computational cost and ensure reliable perception performance, we need lightweight solutions for such models. In this paper, we design a novel structured pruning framework for multi-modal perception networks. Moreover, we proposed a novel structured pruning algorithm Heterogeneous Aware SynFlow (HA-SynFlow), which prunes each modality based on its SynFlow [1] score. We prune the water surface radar-vision fusion model Achelous [2], and results show 24.0% and 7.2% improvement in Frame Per Second (FPS) on GTX1650 and Jetson Orin, respectively, with a 4.4% loss in detection mAP metric. Lastly, our pure radar pruning test shows that radar helps in long-range, occlusion, and difficult scenarios in 2D object detection task.