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 YOLOv 5 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.
维普
万方数据