2D LiDAR is widely used in pedestrian detection due to its low cost and the resistance to external disturbances. However, the disorder and sparsity of the LiDAR points makes pedestrian detection more challenging. In this article, we use a transformation model converting points cloud to images and propose an improved person detection network based on YOLOX. Coordinate Attention (CA) is introduced before pyramidal feature representation to improve the pedestrian feature extraction capability of the model, and the adaptive spatial feature fusion network (ASFF) is added before YoloHead to alleviate the inconsistency across different feature scales. Compared with the YOLOX, the average accuracy (AP) and mean average accuracy (mAP) of the improved network are increased by 1% and 1.4%, respectively, with the average accuracy reaching 95.2%. The single frame point cloud image inference time is 48 ms, which is only 8 ms longer than the original model. The result demonstrates that the proposed model can improve pedestrian detection accuracy and robustness effectively while maintaining real time efficiency.