Pillar3D-Former: A Pillar-Based 3-D Object Detection and Tracking Method for Autonomous Driving Scenes

The 3-D object detection algorithm based on Lidar is one of the fundamental technologies for autonomous vehicle environment perception. At present, compared with the Voxel-based detector, there is still a large gap in the detection accuracy of the pillar-based detector, especially for the detection of small targets. Therefore, in this work, we conduct research around pillar-based detectors and propose a new pillar-based 3-D detector, Pillar3D-Former. We first redesign the pillar feature encoding module and propose global-and-local-enhanced pillar encoding (GLE-PE), which can fully extract the global and local features in the pillar. In addition, a neck called multireceptive-feature aggregation (MRFA) is designed to increase receptive field of model to aggregate more features focusing on localization. Finally, considering that the center-based method is challenging to capture the feature relationship, we naturally combine the center-based method with the transformer to eliminate the inconsistency in feature learning of the traditional two-stage detector. At the same time, we achieve a multiobject tracking algorithm with higher accuracy and robustness based on the detection algorithm by improving the data association and trajectory management strategy. The proposed model is trained and validated on the large-scale dataset nuScenes. The results demonstrate that the proposed Pillar3D-Former achieves 62.89 mean average precision and 69.26 nuScenes detection score (NDS), and the tracking accuracy achieves 0.698 average multiobject tracking accuracy, which is superior to many prevalent mainstream algorithms.