Eyes All Around: Design and Analysis of 360-Degree LiDAR Perception Using Equivariant Feature Learning in Unstructured Traffic

Researchers present a 360-degree LiDAR perception system for autonomous driving that uses rotation equivariant feature learning to handle dense, unstructured urban traffic. Tested on a custom dataset from Indian urban environments, the system achieves strong performance on larger vehicles but struggles with smaller, more variable road users like pedestrians and motorcyclists.

This research addresses a critical gap in autonomous driving perception systems: the transition from limited field-of-view LiDAR processing to full surround 360-degree sensing in chaotic urban environments. Most existing 3D object detectors are developed and validated in structured driving scenarios with standardized road layouts, leaving their behavior under panoramic sensing largely unexplored. The paper's contribution lies in combining sector-wise panoramic processing with rotation equivariant sparse convolutions, a mathematical framework that maintains consistency regardless of sensor orientation changes.

The findings reveal a performance stratification reflecting real-world detection challenges. Cars achieve 92.02% average precision—nearly human-level performance—while buses and trucks reach 80% and 78%, indicating the system handles larger, more predictable vehicles effectively. However, pedestrians (67.45%), motorcyclists (71.20%), and cyclists (73.21%) present significantly harder detection problems, likely due to their smaller scale, variable poses, and unpredictable movement patterns characteristic of Indian traffic.

For the autonomous driving industry, this work highlights both capability and limitations. Equivariant neural networks represent a promising architectural approach for handling the mathematical symmetries inherent in 360-degree sensing, potentially more efficient than brute-force data augmentation. However, the performance gap between vehicle and vulnerable road user detection underscores why Level 4-5 autonomy in dense urban environments remains challenging.

Future research should focus on improving detection for smaller, more dynamic road users through domain-specific data collection, architectural improvements, and potentially sensor fusion strategies that combine LiDAR with complementary modalities.

• →360-degree LiDAR perception with rotation equivariant features achieves 92% accuracy for car detection in unstructured urban traffic
• →Performance degrades significantly for vulnerable road users, with pedestrian detection at 67.45% indicating persistent challenges in dense urban environments
• →Equivariant sparse convolutions offer a mathematically principled approach to handling panoramic sensing without extensive data augmentation
• →Testing on Indian urban traffic reveals substantial real-world complexity beyond structured driving datasets commonly used for validation
• →The performance gap between vehicles (80-92%) and smaller road users (67-73%) identifies key technical hurdles for urban autonomous driving deployment