Point Density-Aware Voxels for LiDAR 3D Object Detection (2203.05662v2)

Abstract: LiDAR has become one of the primary 3D object detection sensors in autonomous driving. However, LiDAR's diverging point pattern with increasing distance results in a non-uniform sampled point cloud ill-suited to discretized volumetric feature extraction. Current methods either rely on voxelized point clouds or use inefficient farthest point sampling to mitigate detrimental effects caused by density variation but largely ignore point density as a feature and its predictable relationship with distance from the LiDAR sensor. Our proposed solution, Point Density-Aware Voxel network (PDV), is an end-to-end two stage LiDAR 3D object detection architecture that is designed to account for these point density variations. PDV efficiently localizes voxel features from the 3D sparse convolution backbone through voxel point centroids. The spatially localized voxel features are then aggregated through a density-aware RoI grid pooling module using kernel density estimation (KDE) and self-attention with point density positional encoding. Finally, we exploit LiDAR's point density to distance relationship to refine our final bounding box confidences. PDV outperforms all state-of-the-art methods on the Waymo Open Dataset and achieves competitive results on the KITTI dataset. We provide a code release for PDV which is available at https://github.com/TRAILab/PDV.

• The paper presents a novel density-aware framework that integrates voxel point centroid localization, KDE-based RoI pooling, and density confidence prediction.
• The methodology improves 3D object detection accuracy, notably boosting vehicle mAPH by +1.25% on the Waymo Open Dataset.
• This advancement enhances detection robustness across varying distances, promising safer autonomous driving technologies.

Point Density-Aware Voxels for LiDAR 3D Object Detection

This paper presents a novel approach to enhancing LiDAR 3D object detection in autonomous driving through the use of Point Density-Aware Voxel networks (PDV). This research arises from the observation that LiDAR point density naturally diminishes with increasing distance due to the angular configuration of the LiDAR sensor, which causes points to diverge. As a result, existing voxel-based methods that treat point density as uniform tend to lose fidelity at greater distances, ultimately hindering accurate 3D object detection. By introducing a density-sensitive metric, PDV aims to address these pitfalls and achieve more reliable object detection across variable distances.

The PDV architecture integrates several key components. The first is Voxel Point Centroid Localization, which improves feature map alignment by using centroid locations of the voxels instead of the voxel centers alone. This approach capitalizes on the raw point cloud data and better captures local geometry, thus refining 3D bounding box proposals without computationally intensive methods such as farthest point sampling (FPS).

Secondly, the Density-Aware RoI Grid Pooling module is introduced. This component utilizes kernel density estimation (KDE) to incorporate local density information into region of interest (RoI) pooling. By using KDE to estimate density and augment RoI features, PDV encodes more granular spatial information that reflects the non-uniformity of point distributions in the point cloud data. This method is further enhanced with a grid point self-attention mechanism that integrates point density positional encoding, allowing the network to focus on differentiating areas with variable density.

The third contribution is the Density Confidence Prediction module, which leverages the relationship between point density and distance by incorporating features like the final bounding box centroid location and the count of LiDAR points within this box into the confidence prediction process. This enhancement allows PDV to provide more informed and reliable bounding box confidence scores, reducing the instance of false positives, particularly at larger distances.

One of the most significant results from the empirical evaluations is that PDV outperforms all current state-of-the-art object detection methods on the Waymo Open Dataset, with improvements noted across vehicle, pedestrian, and cyclist detection tasks at varying distances. Specifically, the vehicle detection task sees an improvement of mAPH by +1.25% at the LEVEL_2 benchmark, highlighting the efficacy of density-aware adjustments. Similarly, performance improvements are noted for smaller and more distant objects like pedestrians and cyclists, proving the method's advantage in high-divergence scenarios.

Practical implications of this work involve advancements in safer and more effective autonomous driving technologies. PDV's enhancement of 3D object detection reliability paves the way for more precise navigation and decision-making capabilities in autonomous systems. Theoretically, this research demonstrates a significant shift towards leveraging intrinsic sensor characteristics more effectively, suggesting that other sensor-based detection frameworks might benefit from density-aware enhancements.

Looking forward, the integration of PDV with other sensor modalities and environments, such as rainy weather affecting point density distributions, might provide further insights into building adaptive, resilient object detection systems. While the primary focus remains on LiDAR, the principles of density-aware voxelization and feature encoding are broadly applicable, offering potential advancements in other areas of 3D spatial analysis and robotics.