Guan Runwei ; Yao Shanliang ; Zhu Xiaohui ; Man Ka Lok ; Lim Eng Gee ; Smith Jeremy ; Yue Yong ; Yue Yutao

Abstract

Current perception models for different tasks usually exist in modular formson Unmanned Surface Vehicles (USVs), which infer extremely slowly in parallelon edge devices, causing the asynchrony between perception results and USVposition, and leading to error decisions of autonomous navigation. Comparedwith Unmanned Ground Vehicles (UGVs), the robust perception of USVs developsrelatively slowly. Moreover, most current multi-task perception models are hugein parameters, slow in inference and not scalable. Oriented on this, we proposeAchelous, a low-cost and fast unified panoptic perception framework forwater-surface perception based on the fusion of a monocular camera and 4DmmWave radar. Achelous can simultaneously perform five tasks, detection andsegmentation of visual targets, drivable-area segmentation, waterlinesegmentation and radar point cloud segmentation. Besides, models in Achelousfamily, with less than around 5 million parameters, achieve about 18 FPS on anNVIDIA Jetson AGX Xavier, 11 FPS faster than HybridNets, and exceed YOLOX-Tinyand Segformer-B0 on our collected dataset about 5 mAP$_{\text{50-95}}$ and 0.7mIoU, especially under situations of adverse weather, dark environments andcamera failure. To our knowledge, Achelous is the first comprehensive panopticperception framework combining vision-level and point-cloud-level tasks forwater-surface perception. To promote the development of the intelligenttransportation community, we release our codes in\url{https://github.com/GuanRunwei/Achelous}.