3D Radar Object Detection Using Deep Learning Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Radar sensors in the automotive industry face challenges in object detection due to their weaker angular resolution, which limits their ability to differentiate objects based on angle, while they can accurately measure radial velocity, resulting in a 3D data cube with a large number of Doppler bins.
Innovation Solution
A computer-implemented method processes 3D radar data comprising range, angle, and Doppler information using deep-learning algorithms to decompose and process the data into three sets of 2D data, applying compression, convolution, dropout, position encoding, and cross-attention algorithms to enhance object detection, ultimately generating a 2D range-angle grid for improved detection capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If radar sensors use traditional signal processing methods to handle 3D data, then they can maintain simple processing algorithms, but the angular resolution remains weak and object detection accuracy is limited
Solution Approach 1:
The patent segments the 3D radar data into multiple 2D planes (range-Doppler, angle-Doppler, range-angle planes) and processes each plane separately using deep learning algorithms. This segmentation allows the system to improve angular resolution through specialized processing on angle-containing planes while keeping the overall processing complexity manageable by dividing the task into smaller, parallelizable components.
Solution Approach 2:
The patent transforms the processing approach by converting 3D radar data into multiple 2D representations, effectively using dimensionality reduction to enable more efficient processing. By representing the data in 2D planes with different dimension combinations (range-angle, range-Doppler, angle-Doppler), the system can apply 2D deep learning algorithms that are computationally more efficient while still capturing the full 3D information through the integration of multiple planes.
2Measurement precision
If radar sensors increase the number of Doppler bins to improve radial velocity measurement, then velocity measurement accuracy is improved, but the data cube becomes larger and processing becomes more complex
Solution Approach 1:
The patent extracts Doppler information specifically from the Doppler dimension in the 3D data cube and processes it separately in 2D planes. By isolating the Doppler processing from the full 3D processing and handling it in dedicated range-Doppler and angle-Doppler planes, the system can maintain high velocity measurement accuracy while reducing the computational burden of processing the complete high-dimensional data cube.
3Measurement precision
If radar sensors process full 3D data cubes with deep learning algorithms, then object detection accuracy is improved, but processing time and computational resources increase significantly
Solution Approach 1:
The patent divides the 3D data processing into multiple 2D plane processing tasks that can be executed in parallel. By segmenting the processing into range-Doppler plane processing, angle-Doppler plane processing, and range-angle plane processing, the system reduces the computational complexity of each individual processing task while maintaining overall detection accuracy through the integration of results from all planes.
Solution Approach 2:
The patent processes only the necessary 2D planes containing relevant information for specific detection tasks rather than processing the entire 3D data cube with full complexity. For example, angle-Doppler planes are processed when angular information is needed, and range-Doppler planes are processed when velocity information is needed, avoiding unnecessary processing of redundant data.
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
This approach effectively processes 3D radar data to improve object detection by enhancing angular resolution and maintaining radial velocity information, making it suitable for embedded systems without increasing processing complexity, thus overcoming the limitations of traditional radar signal processing.
Implementation Method 1
radar technology can measure the Doppler, i.e., the radial speed component, quite accurate
Data Source
AI summary
This disclosure describes systems and techniques for object detection. In aspects, techniques include obtaining 3D data including range data, angle data, and doppler data. The techniques further include processing a deep-learning algorithm on the 3D data to obtain processed 3D data and obtaining processed 2D data from the processed 3D data. The processed 2D data includes range data and angle data.