基于多特征聚合的水面遮挡目标检测算法

Multi feature fusion-based water occlusion object detection algorithm

针对智能船舶在内河航道航行时经常出现的船舶互相遮挡而影响目标检测精度的问题,提出了一种基于多特征聚合的水面遮挡目标检测算法.首先,在骨干网络设置多尺度感受野特征融合结构,融合被遮挡船舶可视区域与周围环境特征;其次,在骨干网络及网络的特征拼接部分添加混合注意力机制,增强网络的长程依赖性,聚合船首和船尾的特征;然后,设计了数据重采样策略,在训练过程中根据船舶类别的数量自适应地调整样本采样频率,缓解数据集中船舶数量的严重不均匀;最后对算法进行验证。结果表明:算法通过聚合被遮挡船舶可视区域与周围环境等多尺度特征,聚合船首、船尾长程特征,相较于原算法精度提升达到了3.3%,有效提升了视觉遮挡状态下水面目标的检测精度.

Aiming at the problem that the object detection accuracy was affected by the mutual occlusion of ships that often occured when intelligent ships navigated in inland waterways,a water surface occlusion object detection algorithm based on multi-feature aggregation was proposed.First,a multi-scale sensory field feature fusion structure was set up in the backbone network to fuse the visible area of the occluded ship with the surrounding environment features.Second,a hybrid attention mechanism was added to the backbone network and the feature splicing part of the network to enhance the long-range dependence of the network,and to aggregate the features of the ship's bow and stern.Then,a data resampling strategy was designed to adaptively adjust the sample frequency according to the number of ship categories during the training process to alleviate the serious unevenness of the number of ships in the dataset.Finally,the algorithm was validated.Results show that the algorithm can effectively improve the detection accuracy of surface targets under visual occlusion by aggregating multi-scale features such as the visible area of the occluded ship and the surrounding environment,and by aggregating the long-range features of the bow and the stern of the ship,with an accuracy increase of 3.3%compared with the original algorithm.