Aiming at the problem of extracting crucial features of occluded pedestrians using convolutional neural network

a high-level semantic feature pyramid for occluded pedestrians detection

HSFPNet

is proposed.The method proposes a global and local crucial feature extraction(GLFE)module

which is based on high-level semantic information

captures long-range dependency relationship and critical marginal information of pedestrians by lightweight multilayer perceptron(MLP)and learnable visual center mechanism

respectively.Meanwhile

a regulated feature pyramid(RFP)with modulation from low-resolution to high-resolution feature layers is introduced.It regulates the high-resolution feature layers to integrate effective features of pedestrians

to obtain more complete and distinctive pedestrian features.On publicly available datasets such as CityPersons and Caltech

specifically on reasonable occlusion subsets

this method achieves an average miss rate of 9.3% and 2.2%

respectively

with a detection speed of 0.16s per image.These results demonstrate the superior performance of this method in addressing pedestrian occlusion.