LiDAR Point Cloud Classification Using 3D Grid Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing obstacle identification methods for autonomous vehicles using lidar-based laser point cloud data are either rule-based, which is difficult to implement for all types of obstacles, or machine learning-based, which struggles with untrained obstacles, leading to inefficiencies in obstacle recognition.
Innovation Solution
A method and apparatus that acquire current pose information of an autonomous vehicle to identify laser point cloud data by utilizing a preset three-dimensional grid map, where N×N map blocks are loaded into a cache, and each laser point data is classified as static or dynamic based on corresponding grid cube types, improving the accuracy of obstacle identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If rule-based methods are used for obstacle identification, then the implementation is straightforward for known obstacle types, but it is difficult to identify all types of obstacles in the real world
Solution Approach 1:
The patent segments the continuous laser point cloud data into discrete grid cubes within a three-dimensional grid map. Each grid cube is independently classified as static or dynamic, enabling systematic processing of obstacle identification while maintaining the ability to handle diverse obstacle types through structured spatial division.
Solution Approach 2:
The patent transitions from two-dimensional map blocks to three-dimensional grid cubes by adding the vertical dimension. This dimensional expansion allows the system to capture and classify obstacles at different heights, significantly improving adaptability to various obstacle types while maintaining manageable complexity through systematic spatial organization.
2Measurement precision
If machine learning-based methods are used for obstacle identification, then the system can learn from training data, but it cannot identify types of obstacles that have not been trained
Solution Approach 1:
The patent creates a universal obstacle identification framework where grid cubes serve as a common representation for all obstacle types. The static/dynamic classification system provides a multi-functional approach that can handle any obstacle type without requiring specific training, as the system universally applies the same spatial classification logic to all detected objects.
Solution Approach 2:
The patent changes the classification parameters from specific obstacle type labels to fundamental spatial properties (static vs. dynamic). This parameter transformation allows the system to maintain high identification accuracy while achieving universal adaptability, as any obstacle can be classified by its motion characteristics rather than requiring pre-trained type-specific models.
3Measurement precision
If a large portion of the three-dimensional grid map is loaded into cache, then the obstacle identification accuracy improves, but the memory consumption increases
Solution Approach 1:
The patent applies local quality by loading only N×N map blocks surrounding the current vehicle position into cache, rather than the entire map. This localized approach ensures high obstacle identification accuracy in the relevant area while significantly reducing memory consumption by excluding distant, irrelevant map regions from the cache.
Solution Approach 2:
The patent performs preliminary loading of the N×N map blocks into cache before obstacle identification is needed. This advance preparation ensures that the required spatial data is readily available in memory when identification is performed, improving accuracy without requiring the entire map to be permanently loaded, thus optimizing memory usage.
Data Source
AI summary
A method and apparatus for identifying laser point cloud data of an autonomous vehicle provided with a lidar. A specific implementation of the method includes: acquiring current pose information of the autonomous vehicle in a world coordinate system in response to receiving a latest frame of laser point cloud data collected by the lidar; acquiring from the cache, based on the current pose information, N×N map blocks centered on map blocks corresponding to the current pose information in a preset three-dimensional grid map and pre-loaded into a cache and are; and executing, for each laser point data in the received laser point cloud data, the laser point data identification operations. The implementation realizes identifying whether each laser point data is a static laser point, and can improve the accuracy rate of identifying a laser point data obstacle.