Aiming at the problem of poor detection accuracy caused by pedestrian crowding and occlusion in natural complex scenes, a dense pedestrian detection algorithm based on YOLOv7 with optimized weights is proposed. First, to address the occluded pedestrian feature extraction problem, the weights of the backbone network are redistributed by the algorithms for typical geometric figures of rectangle and circle. Measuring principles and algorithms of typical plane cross-space efficient multi-scale attention module with cross-spatial learning (EMA), and the correlations between different channel features are learned cross-dimensionally, which can enhance the model’s attention to the visible area of the pedestrian target. Second, to address the problem of high complexity of the detection model, the efficient lightweight connection module (ELCM) is designed to improve the model representation ability and speed up the training speed. Finally, a focused bounding box loss function, Focal-SIoU loss, is constructed, which focuses on suppressing low-quality samples and adds angular loss to improve the detection accuracy of the model. Experimental results demonstrate that the proposed algorithm achieves mean average precisions of 83.7% and 82.6% on the Wider-Person and Crowd Human datasets, respectively, showing significant advantages in dense crowded pedestrian detection tasks.