360-Degree LiDAR system boosts self-driving detection in chaotic Indian traffic
New equivariant learning pipeline achieves 92% car detection accuracy in dense, unstructured urban scenes.
A new research paper titled 'Eyes All Around' presents a 360-degree LiDAR perception pipeline designed for autonomous driving in dense, unstructured urban traffic. The framework combines sector-wise panoramic processing with rotation equivariant sparse convolutions to handle full surround sensing, a departure from typical limited-field-of-view systems. The authors collected a custom dataset using an Ouster OS0 LiDAR across diverse Indian traffic conditions, which feature irregular road layouts, frequent occlusions, and varied road users.
Evaluation shows strong performance for larger vehicles: cars scored 92.02/90.51, buses 80.53/76.34, and trucks 78.59/74.16. However, smaller and more variable road users proved challenging: pedestrians at 67.45/61.02, cyclists at 73.21/69.54, and motorcyclists at 71.20/68.13. The results highlight the difficulty of detecting vulnerable road users in chaotic environments and underscore the need for further refinement in equivariant feature learning for panoramic sensing.
- Uses rotation equivariant sparse convolutions for 360-degree LiDAR perception in unstructured traffic.
- Achieves 92% car detection but only 67% for pedestrians in dense Indian urban scenes.
- Dataset collected using Ouster OS0 LiDAR, covering diverse real-world Indian traffic conditions.
Why It Matters
Advances autonomous driving perception for chaotic, non-standardized roads where current systems fail, especially in developing countries.