A Method with Improved Accuracy and Robustness for Object Detection in Wharf Scenarios

Lidar, radar, and camera are the major candidate sensors used in autonomous boat navigation. Compared with lidar, millimeter-wave radar is more robust under severe weather conditions (e.g., rain, fog). However, the sparsity of radar point cloud results in a false alarm or a missed detection when only radar is used. To improve the accuracy and reliability of object detection under boat docking scenes, a new object detection method is proposed in this paper, which utilizes a newly proposed multi-frame point cloud alignment algorithm to generate a more dense and stable point cloud, and in further a novel clustering algorithm to exploit semantic information of vision to enrich radar points' feature. Moreover, oriented by the requirements of obstacle avoidance in wharf scenarios, a new detection performance metric called Non-Occupation Overlap (NOO) is specifically designed. To demonstrate the proposed method, the in field test is performed using a 77GHz radar and a camera during ship maneuvers in the wharf. Both qualitative analysis and quantitative experiments verify the superiority of the proposed detection method as compared to conventional ones.