Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Been There, Scanned That: Nostalgia-Driven LiDAR Compression for Self-Driving Cars

Ali Khalid, Jaiaid Mobin, Sumanth Rao Appala, Avinash Maurya, Stephany Berrio Perez, M. Mustafa Rafique, Fawad Ahmad

TL;DR

DejaView tackles the challenge of storing and transferring terabytes of autonomous-vehicle LiDAR data by exploiting cross-day redundancy through a cascaded diff against a past reference cloud and a 3D map. The method combines a two-stage diff with a one-way diff, a hybrid coarse/fine nearest-neighbor search, and delta-encoded indices, achieving up to $210\times$ compression with a reconstruction error of $15\text{ cm}$ on two months of data. Key contributions include the cascaded diff framework, efficient diffusion via a hybrid search, and an open-source end-to-end implementation, demonstrated to preserve localization, object detection, and 3D semantic segmentation performance at high compression. This work enables practical offline storage and forensic replay for AV data, reducing network and cloud costs while maintaining actionable fidelity in downstream perception tasks.

Abstract

An autonomous vehicle can generate several terabytes of sensor data per day. A significant portion of this data consists of 3D point clouds produced by depth sensors such as LiDARs. This data must be transferred to cloud storage, where it is utilized for training machine learning models or conducting analyses, such as forensic investigations in the event of an accident. To reduce network and storage costs, this paper introduces DejaView. Although prior work uses interframe redundancies to compress data, DejaView searches for and uses redundancies on larger temporal scales (days and months) for more effective compression. We designed DejaView with the insight that the operating area of autonomous vehicles is limited and that vehicles mostly traverse the same routes daily. Consequently, the 3D data they collect daily is likely similar to the data they have captured in the past. To capture this, the core of DejaView is a diff operation that compactly represents point clouds as delta w.r.t. 3D data from the past. Using two months of LiDAR data, an end-to-end implementation of DejaView can compress point clouds by a factor of 210 at a reconstruction error of only 15 cm.

Been There, Scanned That: Nostalgia-Driven LiDAR Compression for Self-Driving Cars

TL;DR

DejaView tackles the challenge of storing and transferring terabytes of autonomous-vehicle LiDAR data by exploiting cross-day redundancy through a cascaded diff against a past reference cloud and a 3D map. The method combines a two-stage diff with a one-way diff, a hybrid coarse/fine nearest-neighbor search, and delta-encoded indices, achieving up to compression with a reconstruction error of on two months of data. Key contributions include the cascaded diff framework, efficient diffusion via a hybrid search, and an open-source end-to-end implementation, demonstrated to preserve localization, object detection, and 3D semantic segmentation performance at high compression. This work enables practical offline storage and forensic replay for AV data, reducing network and cloud costs while maintaining actionable fidelity in downstream perception tasks.

Abstract

An autonomous vehicle can generate several terabytes of sensor data per day. A significant portion of this data consists of 3D point clouds produced by depth sensors such as LiDARs. This data must be transferred to cloud storage, where it is utilized for training machine learning models or conducting analyses, such as forensic investigations in the event of an accident. To reduce network and storage costs, this paper introduces DejaView. Although prior work uses interframe redundancies to compress data, DejaView searches for and uses redundancies on larger temporal scales (days and months) for more effective compression. We designed DejaView with the insight that the operating area of autonomous vehicles is limited and that vehicles mostly traverse the same routes daily. Consequently, the 3D data they collect daily is likely similar to the data they have captured in the past. To capture this, the core of DejaView is a diff operation that compactly represents point clouds as delta w.r.t. 3D data from the past. Using two months of LiDAR data, an end-to-end implementation of DejaView can compress point clouds by a factor of 210 at a reconstruction error of only 15 cm.

Paper Structure

This paper contains 16 sections, 8 equations, 20 figures, 1 table.

Table of Contents

  1. Introduction
  2. Motivation and Background
  3. DejaView Design
  4. Cascaded diff Operation
  5. Efficient diff Operation
  6. Reference Dataset
  7. Evaluation
  8. Methodology
  9. End-to-end Experiments
  10. Application-level Results
  11. Sensitivity Analysis
  12. Latency and Throughput Evaluations
  13. Comparison with Learning-based Methods
  14. Related Work
  15. Discussion and Future Work
  16. ...and 1 more sections

Figures (20)

  • Figure 1: The source cloud can be reconstructed using the common and exclusive points w.r.t the reference cloud.
  • Figure 2: Common and exclusive points in two point clouds.
  • Figure 3: Compression performance for strawman pipelines.
  • Figure 4: Overview of DejaView. The AV generates source data, which is compressed at edge infrastructure (e.g., at garage or parking lot) to produce a lean representation. This lean representation is transmitted over the network to the cloud for storage. When requested by an application, the cloud decompresses the lean representation to reconstruct the data and delivers it to the application.
  • Figure 5: DejaView's cascaded diff operation uses a three step process to compute a compact representation for the source cloud using a reference cloud and a 3D map. The point clouds with purple outlines constitute the compact representation.
  • ...and 15 more figures