夜间低光照场景下基于多尺度特征融合的密集行人检测算法OA

In complex nighttime scenes,factors such as light variations,pedestrian occlusions,and vary-ing pedestrian scales lead to a decrease in pedestrian detection accuracy.Addressing this challenge,we proposed a dense pedestrian detection algorithm based on multi-scale feature fusion for low-light nighttime scenes.Firstly,a bidirectional path feature aggregation network integrated with self-attention and convolu-tional mixing module was designed.This enhanced multi-scale feature fusion and improved the discrim-inability between pedestrians and background,thus reduced the impact of target scale variations upon de-tection.Secondly,a multi-branch joint detection strategy was constructed.This strategy effectively lever-aged the strong discriminative power of pedestrian contours,legs,and arms in low-light environments to assist detection,reduced the impact of weak light and occlusion.Next,we introduced consideration of cen-ter distances between detection boxes during post-processing.By reducing confidence scores for severely occluded boxes and performing re-selection,we successfully retained correct detections of occluded pedes-trian targets,further decreased the miss detection rate.Finally,the effectiveness of the proposed algo-rithm was verified through experiments.The results show that the log-average miss detection rate is de-creased by 2.8%and 3.3%on the NightOwls and NightSurveillance datasets,respectively.On the LL-VIP dataset,the average precision and recall are improved by 2.1%and 2.9%,respectively.Further-more,all three proposed improvement modules contribute to enhancing the detection performance of the al-gorithm.