swc-17/SparseDrive

What it solves

SparseDrive addresses the complexity and inefficiency of end-to-end autonomous driving systems. It aims to reduce the high training and inference costs associated with traditional dense scene representations while improving safety-critical metrics, specifically the collision rate during motion planning.

How it works

SparseDrive utilizes a "Sparse-Centric" paradigm that unifies detection, tracking, and online mapping into a single symmetric sparse perception structure. The process follows these steps:

1. Feature Encoding: Multi-view images are encoded into feature maps.
2. Sparse Perception: The system learns a sparse scene representation of the environment.
3. Parallel Motion Planning: Instead of a sequential pipeline, it performs motion prediction and planning simultaneously. This is supported by a hierarchical planning selection strategy and a collision-aware rescore module to ensure safe trajectories.
4. Temporal Modeling: An instance memory queue is used to handle temporal data.

Who it’s for

This project is designed for researchers and engineers working on autonomous driving, specifically those focusing on end-to-end perception-to-planning pipelines and efficient scene representation.

Highlights

• High Efficiency: Significantly reduces training time (e.g., from 144 hours to 20-30 hours) and increases FPS compared to previous SOTA methods like UniAD.
• Superior Safety: Achieves a lower collision rate on the nuScenes benchmark.
• Unified Architecture: Combines multiple driving tasks into one sparse instance representation.
• Parallel Design: Optimizes motion planning by treating it as a parallel process to motion prediction.