基于YOLO v3的落水人员检测

Detection of People Falling into Water Based on YOLO v3

针对落水人员所处水域的复杂性以及波纹、阳光等因素导致对落水人员检测的准确率较低,以及在检测小目标时,经典检测算法易出现误检漏检情况,提出一种改进的YOLO v3目标检测算法。使用k-means++聚类算法对自有落水人员数据集进行聚类,得到更适合落水人员的锚框,从而提高检测速度与精度;在网络中加入通道注意力机制模块,其关注通道信息,可以学习到不同通道特征的重要程度,根据重要程度为每个通道分配相应的权重,从而让网络关注重要的特征,抑制不重要的特征,提高重要特征的表征能力;引入感受野模块(RFB)来增大浅层特征图的感受野,从而提高小目标检测精度。最后,在自制的落水人员数据集上对该算法进行了验证,结果表明,该算法在检测效果上优于原始YOLO v3。

In view of the complexity of the water area where the person falling into the water and low accuracy of the detection of the person falling into the water caused by the ripples,sunlight and other factors,as well as false detection and missed detection of the classic detection algorithm when detecting small targets,we propose an improved YOLO v3 target detection algorithm.We use the k-means++clustering algorithm to perform clustering analysis on the data set of people falling into the water to obtain an anchor box that is more suitable for people,thereby improving the detection speed and accuracy.A channel attention mechanism module,which is added in the network and focuses on channel information,can learn the importance of different channel features and assign different weights for each channel based on the importance,so that the network pays attention to important features,suppresses unimportant features,and improves the characterization ability of important features.We introduce Receptive Field Module(RFB)to increase the receptive field of the shallow feature map,thereby improving the accuracy of small target detection.Finally,the algorithm is verified on a self-made data set of people falling into the water.It is showed that the proposed algorithm is superior to the original YOLO v3 in terms of detection effect.