Lidar Intensity Map Localization Without Sensor Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional autonomous vehicle localization techniques face challenges in achieving accurate, real-time localization, especially in geometrically non-distinctive or repetitive environments, and require laborious calibration of Lidar intensity returns, which limits their practicality and effectiveness.
Innovation Solution
A deep neural network-based localization system that embeds Lidar intensity maps and online Lidar sweeps in a common space, allowing for vehicle localization without the need for complicated calibration, and is effective across various environments, including highways and urban areas, using GPU implementations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If intensity-based localization methods are used, then localization precision can be improved, but the system requires laborious calibration of Lidar intensity returns
Solution Approach 1:
The system performs self-calibration by automatically adapting to intensity characteristics of different Lidar sensors through the neural network training process. The network learns sensor-specific intensity patterns during training and generalizes to new sensors without manual calibration, allowing the system to calibrate itself autonomously.
Solution Approach 2:
The system changes the approach from fixed intensity thresholds to learned intensity representations. The neural network transforms raw intensity values into calibrated feature representations that are adaptive to different sensors and environmental conditions, eliminating the need for manual parameter tuning.
2Measurement precision
If geometric localization methods are used, then high-precision localization can be achieved, but the system fails in geometrically non-distinctive or repetitive environments
Solution Approach 1:
The system combines geometric features (from point cloud structure) with intensity features (from Lidar reflectivity) to create a composite representation. This composite approach leverages the strengths of both methods: geometric precision and intensity-based environmental distinctiveness, enabling robust localization in repetitive environments where geometry alone fails.
Solution Approach 2:
The neural network merges multiple feature types including geometric structure, intensity values, and semantic information into a unified localization framework. This integration allows the system to disambiguate locations in repetitive environments by combining subtle intensity variations with geometric context.
3Adaptability or versatility
If image-based localization methods are used, then robust localization can be achieved, but the localization precision is lower compared to geometric methods
Solution Approach 1:
The system replaces traditional image-based processing with a neural network-based approach that directly processes raw Lidar intensity data. This substitution enables the system to achieve both the robustness of image-based methods (through learned features) and the precision of geometric methods (through direct Lidar measurement utilization).
4Measurement precision
If Lidar intensity calibration is performed manually, then accurate intensity returns can be achieved, but the calibration process is laborious and prevents practical deployment
Solution Approach 1:
The system performs self-calibration by automatically adapting to intensity characteristics of different Lidar sensors through the neural network training process. The network learns sensor-specific intensity patterns during training and generalizes to new sensors without manual calibration, allowing the system to calibrate itself autonomously.
Solution Approach 2:
The system replaces expensive, time-consuming manual calibration procedures with a computationally efficient neural network approach. The calibration is performed through standard training data collection and processing, eliminating the need for specialized calibration equipment and expert intervention.
5Measurement precision
If maps are re-captured for each sensor change, then accurate localization can be maintained, but the process becomes impractical for frequent sensor updates
Solution Approach 1:
The neural network model is designed to be sensor-agnostic and can adapt to different Lidar sensors through training on data from each sensor type. This universality allows the same localization system to work with multiple sensor generations and models without requiring separate map recapture processes, making the system universally applicable across sensor updates.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The solution provides improved localization precision, robustness, and the ability to work with uncalibrated data, generalizing across different Lidar sensors and handling dynamic objects, while reducing the need for frequent recalibration and map recapture.
Implementation Method 1
The Lidar unit also includes circuitry to measure the time of flight (ToF)—i.e., the elapsed time from emitting the light signal to detecting the return signal. The time of flight is used to determine the distance of the Lidar unit to the detected object.
Implementation Method 2
During operation, each channel emits a light signal into the environment that is reflected off of the surrounding environment back to the receiver.
Data Source
AI summary
Aspects of the present disclosure involve systems, methods, and devices for autonomous vehicle localization using a Lidar intensity map. A system is configured to generate a map embedding using a first neural network and to generate an online Lidar intensity embedding using a second neural network. The map embedding is based on input map data comprising a Lidar intensity map, and the Lidar sweep embedding is based on online Lidar sweep data. The system is further configured to generate multiple pose candidates based on the online Lidar intensity embedding and compute a three-dimensional (3D) score map comprising a match score for each pose candidate that indicates a similarity between the pose candidate and the map embedding. The system is further configured to determine a pose of a vehicle based on the 3D score map and to control one or more operations of the vehicle based on the determined pose.