Semantic Segmentation Clustering for Radar Point Cloud Object Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional clustering techniques for autonomous vehicle perception systems struggle to accurately identify objects in close proximity, such as pedestrians and vehicles, due to partial occlusion and overlapping fields of view, leading to incorrect clustering of radar and LiDAR point cloud data.
Innovation Solution
The use of semantic labels generated using machine-learning techniques is introduced to improve clustering resolution, where these labels are associated with each point cloud data point, allowing for adaptation of clustering parameters based on object type, such as increasing or decreasing radial distance parameters, and incorporating localization information to enhance the accuracy of distance-based clustering algorithms like DBSCAN.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional clustering techniques are used for point cloud data, then the clustering process is simple and fast, but the object identification accuracy deteriorates when objects are in close proximity
Solution Approach 1:
The patent applies preliminary action by generating semantic labels for point cloud data before performing clustering. These pre-computed semantic labels provide contextual information about objects (such as vehicle types, pedestrians, cyclists) that guides the subsequent clustering process, enabling more accurate separation of closely positioned objects without requiring complex clustering algorithms during the main processing stage.
Solution Approach 2:
The patent changes clustering parameters dynamically based on semantic labels. Different cluster radius parameters are selected depending on the semantic category of objects detected - for example, larger radii for vehicle clusters and smaller radii for pedestrian clusters. This parameter adaptation allows the clustering process to achieve higher accuracy for different object types while maintaining computational efficiency.
2Measurement precision
If clustering radius is increased to capture more points, then clustering completeness improves, but the ability to distinguish close objects deteriorates
Solution Approach 1:
The patent applies local quality by using different cluster radius parameters for different spatial locations and object types. Instead of using a uniform cluster radius across the entire point cloud, the system selects specific radius values based on the semantic labels of detected objects - for example, using smaller radii for pedestrians and cyclists who occupy less space, and larger radii for vehicles. This localized parameter adaptation enables accurate object separation while maintaining complete point cloud coverage for each object type.
3Measurement precision
If semantic labels are incorporated into clustering, then object identification accuracy improves, but processing time increases
Solution Approach 1:
The patent minimizes processing time impact by performing semantic label generation as a preliminary step before clustering. The semantic segmentation model processes the point cloud data once to generate labels that are then reused during the clustering stage. This approach avoids the need for iterative or repeated semantic analysis during clustering, thereby maintaining computational efficiency while achieving high accuracy through label-informed parameter selection.
Data Source
AI summary
The disclosed technology provides solutions for improving perception systems and in particular for improving object identification based on sensor point cloud data, such as radar point cloud data. A process of the disclosed technology can include steps for receiving point cloud data comprising a plurality of radar points, wherein each of the radar points corresponds with a first object or a second object in an environment, generating a semantic label for each of the radar points, and clustering the plurality of radar points based on the semantic label for each of the radar points, to generate a first point cloud cluster for the first object and a second point cloud cluster for the second object. Systems and machine-readable media are also provided.