HD Map BEV Fusion for Accurate 3D Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle systems face challenges in accurately detecting 3D objects due to the limitations of current sensor data fusion methods, leading to suboptimal navigation and decision-making capabilities.
Innovation Solution
A high-definition map fusion technique is employed, combining raster maps, vector maps, and point cloud data to generate bird's eye view representations, which are then fused into a single image for enhanced object detection using neural networks, improving the accuracy of 3D object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional sensor data fusion methods are used for 3D object detection, then the system complexity is reduced, but the detection accuracy deteriorates
Solution Approach 1:
The patent segments the complex data fusion process into distinct modules: a raster map processing module that generates 2D feature maps, a vector map processing module that creates geometric representations, and a point cloud processing module that generates 3D spatial information. Each module processes specific data types independently before fusion, reducing overall system complexity while maintaining high detection accuracy through specialized processing pipelines.
Solution Approach 2:
The patent introduces an intermediary bird's-eye view (BEV) representation as a common framework for fusing raster maps, vector maps, and point clouds. This BEV space acts as a mediator that transforms diverse data sources into a unified coordinate system, enabling accurate 3D object detection without requiring direct complex interactions between all data types, thus simplifying the fusion architecture.
2Measurement precision
If multiple data sources (raster maps, vector maps, point clouds) are integrated, then object detection accuracy is improved, but the processing time increases
Solution Approach 1:
The patent performs preliminary processing of raster maps and vector maps to generate pre-computed 2D feature maps and geometric representations before the actual object detection task. These pre-processed intermediate results are stored and readily available when point cloud data arrives, enabling rapid fusion without reprocessing the entire dataset, thus reducing real-time processing time while maintaining accuracy.
Solution Approach 2:
The patent transforms 3D point cloud data into a 2D bird's-eye view representation, matching the dimensionality of raster and vector map features. This dimensional transformation enables efficient element-wise fusion operations between different data sources without requiring computationally intensive 3D-3D or 3D-2D alignment operations, significantly reducing processing time while preserving detection accuracy.
Data Source
AI summary
Provided are methods for high definition map fusions for 3D object detection. Some methods described also include obtaining, with at least one processor, raster maps, vector maps, and point cloud data and extracting, with the at least one processor, features from the raster maps, vector maps, and point cloud data to generate respective bird's eye view (BEV) representations. The methods also include fusing, with the at least one processor, the BEV representation of the raster map features, the BEV representation of the vector map features, and the BEV representation of the point cloud features into a fused BEV image. Additionally, the methods include detecting, with the at least one processor, objects in the fused BEV image. Systems and computer program products are also provided.