Pedestrian Detection Algorithm Based on SE-Attention Block

Pedestrian detection is an important and challenge task.Pedestrian detection is the most fundamental problems in the practical application of computer vision.Besides having independent value in video surveillance and autonomous driving, it is also the basis of computer vision tasks such as pedestrian re-identification,person retrieval and person tracking. This paper proposed that the transformer with the Squeeze Excitation(SE)-attention block added to the backbone is learned to obtain significance regions and input into a pedestrian detection model for detect person. Compared with the classic pedestrian detection algorithms, it can preliminarily estimate the approximate position of the target object more efficiently,form the significance areas of the object detection,and achieve the purpose of rapid target positioning.The method in this paper achieves an mAP of 48.3% on the CityPerson dataset, and detects one image every 0.35 seconds. Experiments demonstrate that the approach based on SE-attention block can effectively improve the pedestrian detection efficiency in scenarios with insufficient lighting, occlusions, and other challenging conditions.