# Parallel Neural Computing for Scene Understanding from LiDAR Perception in Autonomous Racing

**Authors:** Suwesh Prasad Sah  
**arXiv:** [2412.18165](https://arxiv.org/abs/2412.18165)  
**Date:** 2024-12-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2412.18165v1/

## Key points

- Autonomous driving in high-speed racing environments poses challenges for scene understanding due to rapid changes in the track environment
- Traditional sequential network approaches struggle to keep up with real-time knowledge and decision-making demands
- The Parallel Perception Network (PPN) architecture consists of two independent neural networks - segmentation and reconstruction networks - running in parallel on separate accelerated hardware
- PPN utilizes true hardware-enabled parallelism to match the high velocity of the autonomous agent, resulting in faster processing speeds
- The model is trained on a system equipped with two NVIDIA T4 GPUs using various loss functions, including edge preservation, leading to a 2x speedup in model inference time compared to traditional configurations
- GPUs are utilized as hardware accelerators for high-performance computing during both training and inference stages, leveraging advancements in GPU computing and performance within NVIDIA's CUDA platform
- By employing separate accelerated hardware for true hardware-enabled parallel computing, latency issues in real-time perception for multiple tasks in autonomous driving scenarios are effectively mitigated

## Summary

, , , , 

Autonomous driving in high-speed racing environments presents unique challenges for scene understanding. The rapid changes in the track environment make it difficult for traditional sequential network approaches to keep up with real-time knowledge and decision-making demands. To address this issue, a novel baseline architecture called the Parallel Perception Network (PPN) has been proposed. This model consists of two independent neural networks - segmentation and reconstruction networks - running in parallel on separate accelerated hardware. By utilizing true hardware-enabled parallelism, the PPN can match the high velocity of the autonomous agent, resulting in faster processing speeds. The PPN takes raw 3D point cloud data from LiDAR sensors as input and converts it into a 2D Bird's Eye View Map on both devices. Each network within the PPN independently extracts input features along space and time dimensions, producing outputs simultaneously. The model is trained on a system equipped with two NVIDIA T4 GPUs using various loss functions, including edge preservation. This training methodology results in a 2x speedup in model inference time compared to traditional sequential configurations. Furthermore, GPUs are utilized as hardware accelerators for high-performance computing during both training and inference stages of deep neural networks. This utilization leverages advancements in GPU computing and performance, particularly within NVIDIA's CUDA platform. By employing separate accelerated hardware for true hardware-enabled parallel computing, latency issues in real-time perception for multiple tasks in autonomous driving scenarios are effectively mitigated. In conclusion, the Parallel Perception Network represents a significant advancement in scene understanding from LiDAR perception in autonomous racing contexts. By combining sophisticated models with cutting-edge hardware acceleration techniques, this approach demonstrates enhanced efficiency and effectiveness in processing complex environmental data at high speeds.