联合ACF与YOLOv3的目标识别方法研究

Research on Target Recognition Method Based on ACF and YOLOv3

为提升YOLOv3算法对小目标的检测性能,引入基于聚合信道特征的区域提案方法,提取潜在区域送至YOLOv3网络进行检测;以soft-NMS算法改进NMS算法,减少检测框误删以及目标漏检的几率,提升了模型的检测精确率。实验结果表明,相比于传统YOLOv3算法,研究改进的YOLOv3算法在弹库人员检测方面的性能更优。

In order to get a better performance of YOLOv3 algorithm for small targets,this paper used a region proposal method based on the aggregated channel feature(ACF),extracted potential areas and sends them to the YOLOv3 network for detection.As a result,the soft-NMS algorithm improved the NMS algorithm,reduced the probability of false deletion of the bounding box and the missed detection of the target,and increased the detection accuracy of the model.Experimental results show that the improved YOLOv3 algorithm has better performance in the detection of pedestrian in missile warehouse compared with the traditional YOLOv3 algorithm.