【摘要】 为快速、准确地检测船舶目标,提出一种基于YOLOv3算法的船舶双目视觉检测与定位方法。在特征学习时针对样本中不同船舶长宽比例,重新聚类样本中心锚点框,增强对船舶检测的准确性;利用SURF算法进行特征匹配,并引入双目测距算法,实现目标的测距与定位。实验结果表明,该方法在每秒传输图片30帧的情况下,平均检测精度达到94%,在1 n mile内的目标平均定位误差为11 m左右,与现有检测算法相比,具有更好的实时性、准确性。该方法对智能船舶视觉感知信息与雷达、AIS信息的融合,以及避碰辅助决策具有非常重要的作用。更多还原

【Abstract】 In order to detect target ships quickly and accurately, a binocular vision detection and positioning method for ships based on YOLOv3 algorithm is proposed. During feature learning, the center anchor frames of the samples are re-clustered for the different ship length-width ratios in the samples to enhance the accuracy of ship detection; the feature matching is carried out by the SURF algorithm, and the binocular ranging algorithm is introduced to achieve target ranging and positioning. The experimental results show that, under the condition of 30 frames per second, the average detection accuracy of this method is 94%, and the average positioning error of the targets within 1 n mile is about 11 m. Compared with the existing detection algorithms, it is of better real-time performance and accuracy. It plays a very important role in the information fusion from the visual perception, radars and AIS, as well as collision avoidance auxiliary decision for intelligent ships.更多还原