Light detection and ringing (LiDAR) point clouds exhibit significant discreteness, with information density much lower than that of images, resulting in significant challenges for pedestrian detection in this background and exhibiting low detection accuracy. To address the above issues, a LiDAR pedestrian detection method was proposed based on improved complex you only look once (complex-YOLO). Firstly, a new network feature extraction structure was designed to enhance the feature learning ability of the backbone network, in order to more fully mine information in situations of data scarcity. In addition, the designed cascaded feature pyramid network and its multi branch linear fusion method can fuse features of different scales and depths, improve the generalization performance of the network, and cope with data feature distortion. During the training process, enhanced intersection over union (EIoU) was adopted to accelerate the convergence speed of the network. Through experimental verification on the dataset, the detection accuracy of the improved algorithm has been improved by 3.03%. This study is helpful for improving pedestrian detection accuracy in sparse data situations.