ESFormer: A Pillar-Based Object Detection Method Based on Point Cloud Expansion Sampling and Optimised Swin Transformer

In recent years, with the wide application of autonomous driving, surveillance, and robotics, the demand for accurate object detection in efficient object scenarios has surged. However, traditional object detection methods often face the challenge of difficulty in balancing detection accuracy and processing speed when dealing with the dynamic characteristics of fast-moving objects. To this end, this study proposes an innovative target detection method based on Pillar point cloud, by designing a point cloud enhancement module called the Point Cloud Expansion Sampling Tool (PCES-Tool) and combining it with a Swin Transformer architecture based on optimized patch partition (Swin-T+). to solve the above problems. By adopting the inflated point cloud approach, the method enhances the spatial representation of the object to achieve faster object detection. And Swin Transformer further enhances the model’s ability to capture multi-scale features through its hierarchical design, enabling it to recognize objects at different scales and distances. The method is evaluated on the KITTI, NuScenes and WOD standard datasets, and the results show that significant improvements are achieved in the detection accuracy with both three-dimensional intersection and merger ratio (3D IoU) and two-dimensional intersection and merger ratio (2D IoU). The experimental results prove that the method performs well in high-speed detection scenarios, indicating that the combination of the PCES-Tool and Swin Transformer can effectively solve the challenges in point cloud object detection. This study provides important technical support for the future development of real-time object detection systems.