密集行人检测方法研究

Research on dense pedestrian detection method

针对现有行人检测算法面对遮挡、尺度不一等问题表现出来漏检率和误检率高的情况,提出一种基于改进YOLOv5的密集行人检测方法Improved-YOLOv5.采用改进BIFPN网络替换原有PANet,增强特征融合网络对于特征信息的利用率和对于小尺度行人的关注度.采用EIoU Loss替换原有CIoU Loss,提高模型的回归精度和收敛速度.提出一种新的后处理算法T-NMS,通过增加一个额外的阈值,提高模型对于密集场景下行人密度的区分能力,并在模型开销增加不大的前提下降低了漏检率.实验结果表明,在Citypersons数据集上,所提密集行人检测方法Improved-YOLOv5相比原YOLOv5算法在不同程度遮挡的子集上检测效果均有明显提升,尤其是高遮挡Heavy子集的MR-2降低了4.2%,达到53.1%,表明改进方法在密集行人检测中具有较好的性能.

Aiming at the situation that existing pedestrian detection algorithms show high missed and false detection rates in the face of problems such as occlusion and different scales,an Improved pedestrian detection method based on improved YOLOv5 was proposed.The improved BIFPN network was used to replace the original PANet to enhance the feature fusion network s utilization rate of feature information and its attention to small-scale pedestrians.The original CIoU Loss was replaced by EIoU Loss to improve the regression accuracy and convergence speed of the model.A new post-processing algorithm T-NMS was proposed.By adding an extra threshold,the model can improve the ability to distinguish the density of pedestrians in dense scenes,and reduce the missing rate under the premise of a small increase in model overhead.The experimental results showed that compared with the original YOLOv5 algorithm in Citypersons dataset,the improved pedestrian detection method has improved significantly in detecting subsets with different degree of occlusion.Especially,the detection effect of Heavy subset with high occlusion was reduced by 4.2%to 53.1%.It was shown that the improved method has better performance in dense pedestrian detection.