Pillar-MTNet: A Point Cloud Vehicle Target Detection Algorithm Based on Pillar and Swin Transformer

Abstract Object detection in point clouds is a crucial component of numerous robotics applications, including autonomous driving. Developing an efficient yet deployable backbone network capable of extracting feature is a foundational challenge within the domain of 3D perception. In contrast to convolution, the attention mechanism employed in Transformers is better suited for effectively modeling long-range relationships and is more readily deployable in real-world applications. In this paper, we designed a new point cloud vehicle target detection algorithm called Pillar-MTNet for better detection accuracy. We introduced Mobile Block into the network to replace the convolution operation of the traditional backbone network to reduce the network parameters and improve the feature extraction capability of the network. Additionally, for the purpose of extracting more profound semantic features, we also incorporate Swin Transformer Block (STB) to process the feature map subsequent to the final down-sampling and Mobile Block. RDIoU is adopted to optimize detection results and use it to guide bounding box regression and supervise classification. Extensive comparative experimentations on the classic Open Dataset Kitti and Waymo shows that Pillar-MTNet outperforms previous classical methods in vehicle target detection accuracy.