Improved SENet for Pedestrian Detection

This paper presents an enhanced version of the conventional SENet algorithm in computer vision to tackle the problem of missed or challenging detection of pedestrians when they are obscured. The proposed approach introduces a new method for enhancing the feature extraction performance of pedestrian targets. The paper employs an attention module to learn the correlation between the spatial information of feature channels and feature maps. Next, a distance cross-merge ratio loss function is introduced to improve the regression of detection frames by focusing on the cross-merge ratio between candidate frames and real frames. Last, a non-maximum suppression algorithm is employed for post-processing to eliminate redundant prediction frames and retain more accurate prediction frames, ultimately leading to the best object detection location. Experimental results indicate that the proposed network has a simpler structure and achieves higher detection accuracy compared to traditional methods when tested on the WiderPerson dataset.