Casgraph: A Cascade Attention Graph Network Using Kernel Density Estimation Based on LiDAR Point Clouds for 3D Object Detection

Engineering notes

Extensive experiments on the KITTI dataset have illustrated the effectiveness of our Casgraph framework which can obtain more promising results than other state-of-the-art algorithms.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

3D object detection based on LiDAR point clouds has recently attracted considerable attention, particularly in fields such as autonomous driving. However, LiDAR often collects a smaller amount of point cloud data for distant and small objects, leading to the loss of fine-grained details of the objects. Additionally, existing point-based object detection suffers from the drawbacks of relying on point sampling quality. Moreover, traditional cascade structure design results in a training imbalance of positive and negative samples between previous and current stages. To tackle these problems, we propose a novel 3D detection framework called Casgraph, which utilizes dynamic farthest voxel sampling to select representative points within proposed boxes. Subsequently, we employ centroid-based Kernel Density Estimation to obtain point density features which helps capture fine-grained features of objects. The overall features of sampled points are then aggregated and propagated using a k-nearest neighbor-based graph neural network to capture the geometric features of the objects. Finally, the output features are further fed into a multi-stage cascaded network using the multi-head attention mechanism to balance the proposal quality from different stages. Extensive experiments on the KITTI dataset have illustrated the effectiveness of our Casgraph framework which can obtain more promising results than other state-of-the-art algorithms.

Links and sources

• Official / arXiv page

Comments