Efficient dense pedestrian detection based on transformer

Aiming at the problems of occlusion between pedestrians and small-scale pedestrians, as well as the huge number of model parameters and computation for dense pedestrian detection by Deformable DETR. In this paper, an efficient algorithm for dense pedestrian detection is proposed. The algorithm firstly uses EfficientNet network to replace ResNet50 network to make the model lightweight. Secondly, this paper proposes a cross-fertilization module that fuses spatial information into multi-scale features and fuses the features of each scale several times, in order to obtain more occluded and small-scale pedestrian feature information to improve the problem of feature loss caused by network occlusion. At the same time, this paper introduces noise removal training, so that the improved detection algorithm can find and locate dense pedestrians faster. In this paper, the improved algorithm is experimentally evaluated on the CrowdHuman dataset, and the detection accuracy is improved by 1.9 percentage points compared to the baseline model with the number of parameters and floating-point computation almost halved.