Infrared Pedestrian Detection Method Based on Attention Model
Wenxiu Wang · Pattern Recognition and Image Analysis · 2021
Abstract A new infrared pedestrian detection method based on semantic segmentation is proposed. First, an attention model is introduced using traditional semantics to enhance the attention area of the pedestrian target. Then, the feature maps before and after enhancement are analyzed, and the multiscale effect is improved. The scale-pooling module operates on the final layer of the convolution output to concatenate the results. Finally, 10 000 independently developed infrared pedestrian images with multiple transformations are used to create semantically segmented datasets. Experimental results show that, compared with traditional methods, the proposed semantic segmentation network model’s detail and multiscale performance is better. The mean intersection-over-union is improved by 5.51%, and the pixel accuracy is improved by 4.75%. This method can be effectively applied to improve infrared pedestrian target detection.