基于概率图与视觉显著性的海面目标检测方法

Maritime object detection method based on probabilistic graph and visual saliency

针对基于视觉的传统海面目标检测算法在水面无人艇的自动避碰应用中存在检测精确率、召回率低以及对复杂场景的适应性不足的问题,提出一种基于概率图与视觉显著性的海面目标检测算法。首先利用概率图模型分割出原始图像中的海界限区域与海面孤立目标;然后针对海界限区域子图像特点,设计了一种基于方向抑制的梯度特征,并结合背景先验改进频率调谐显著图,利用特征融合的方法提取海界限区域的潜在目标。实验结果表明,该算法能够有效抑制云、飞鸟、海天线和海杂波的背景干扰。与传统方法相比,提出的方法具有更高的精确率与召回率,且满足无人艇自动避碰实时性的要求。

Aiming at the problems that conventional maritime object detection methods based on vision have poor performance in detection precision rate,recall rate and insufficient adaptability to complex scenes in the automatic collision avoidance application of unmanned surface vehicle(USV),this paper proposed a new maritime object detection algorithm based on probabilistic graph and visual saliency.Firstly,the algorithm used a probabilistic graphical model to extract the sea-sky-junction area and isolated objects of sea-surface area in the original image.Then,according to the characteristics of the sub-image of sea-sky-junction area,the algorithm extracted the potential objects of the sea-sky-junction area by combining a designed gradient feature based on directional suppression with an improved frequency-tuned saliency feature based on background prior.Experimental results show that the proposed algorithm can effectively suppress the background interferences such as cloud,bird,sea-sky line and sea clutter.Compared with conventional methods,the proposed algorithm has higher precision and recall rate and can also meet the real-time requirement of automatic collision avoidance for USV.