全域海面障碍检测研究

As a key technique for the platforms on the sea, automatic obstacle detection on sea surface should be carried out correctly and steadily to ensure the platforms avoid obstacles safely and automatically. However, due to changeable and complicated conditions on the sea, diverse and extensively distributed obstacles, and adverse working and sea situation of obstacle detection sensors, the obstacles detected directly by those sensors can’t meet the needs of automatic obstacle avoidance. In this proposal, by dealing with visible optical image and radar dots obtained on the sea surface, obstacle preprocessing and detection algorithms are studied to detect obstacles of all range. We will focus on problems of illumination and spray on the sea for images, large false alarms and low detection probability for radar dots. By analyzing the distribution properties, a segmentation model will be proposed, which is suitable for extracting sky and sea regions. Then, statistical features characterizing illumination areas and spray on the sea will be studied. Regarding features of sky ande sea regions, and those of illumination areas and spray on the sea as globally intrinsic cues and locally extrinsic cues respectively, multi-obstacles will be detected based on salient map. In terms of radar dots, the issues of large false alarms and low detection probability will be studied. An obstacle detection model adaptive to dense clutter environment will be built based on random finite set by analyzing properties of obstacle and false alarm dots from space and time dimensions to detect distant obstacles effectively. Those achievements will be validated on our platforms.

准确稳定的海面障碍自动检测是海上平台安全自主避障的前提，也是其关键技术之一。由于海面情况复杂多变、障碍多种多样且分布广、障碍探测传感器工况和海况恶劣，传感器的探测结果无法满足自主避障算法要求。本课题以海面可见光图像和雷达点迹为处理对象，研究障碍预处理和检测算法，重点解决图像中海面光照及浪花和雷达点迹中检测概率低及虚警多引起的障碍检测问题，实现全域海面障碍检测。针对海面光照和浪花问题，从海面图像中场景分布属性出发，建立适于提取天空和海面区域的分割模型。然后统计海面光照和浪花区域特征，将天空和海面区域特征作为内在全局线索；海面光照和浪花区域特征作为外在局部线索，实现海面多个障碍显著性检测。针对雷达点迹检测概率低和虚警多问题，从障碍和虚假雷达点迹在空域和时域物理表现特性出发，基于随机有限集方法，建立适于在强杂波环境下空域粗-精关联的高斯混合概率密度海面障碍检测模型，实现高效稳定远距障碍检测。

无人艇是无人作战系统的重要跨域节点，是改变未来海上游戏规则的颠覆性技术，将颠覆传统海战样式，催生全新海洋装备体系。准确稳定的海面障碍自动检测是无人艇自主避障的前提，也是其关键技术之一。由于海面情况复杂多变、障碍多种多样且分布广、障碍探测传感器工况和海况恶劣，传感器的探测结果无法满足自主避障算法要求。本课题以海面可见光图像和雷达点迹为处理对象，研究障碍预处理和检测算法。基于此课题从2.5维激光点云目标检测和扩展目标跟踪，S4-SLAM激光点云港口建模，基于定位-分类序列的图像目标检测，时空上下文融合的无人艇海面目标跟踪和研究成果整体验证几个方面开展研究。S4-SLAM激光点云港口建模采用业内通用的KITTI数据集进行了多个场景的测试和验证，可达到5Hz定位输出/1Hz 高精度地图输出的实时定位和建图性能，且定位误差<1%。结果已上传至KITTI网站，最高排名前15，相关成果发表于机器人顶刊Autonomous Robots。同时利用此研究成果在青岛某港口进行建图实验，通过对所构建地图进行精度分析，在总长度为1.8公里的港口区域，其建图误差在30cm范围内。基于点云数据的目标检测跟踪和图像数据的目标检测和跟踪方法应用于2019年海军小型无人艇竞标中，获得第一名。