1D Radar Elevation Ambiguity via CNN Depth Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current radar systems with one-dimensional horizontal antenna arrays lack resolution in elevation, making it difficult to accurately detect and differentiate objects based on their height, which is crucial for advanced driver assistance systems and autonomous driving applications.
Innovation Solution
The use of an encoder-decoder structured deep convolutional neural network (CNN) to process radar frames and predict target elevations by assigning depths to azimuth-elevation pairs, leveraging a dataset of compensated ground truth depth maps acquired from 2D range sensors like LiDAR or infrared sensors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a one-dimensional horizontal antenna array is used, then the device complexity is reduced and manufacturing is easier, but the measurement precision in elevation is lost
Solution Approach 1:
A deep convolutional neural network (CNN) is introduced as an intermediary between the simple 1-D radar array and the complex task of elevation resolution. The CNN processes radar return data and predicts depth maps that include elevation information, effectively mediating between the limited hardware capabilities and the requirement for precise 3-D object detection and differentiation.
Solution Approach 2:
The patent replaces the mechanical approach of using a complex 2-D or 3-D antenna array with a computational approach using deep learning. Instead of physically complicating the radar hardware to achieve elevation resolution, the system uses software-based neural networks to process simple 1-D array data and extract elevation information through trained models.
2Measurement precision
If a two-dimensional or three-dimensional antenna array is used, then the measurement precision in elevation is improved, but the device complexity increases
Solution Approach 1:
The patent substitutes complex mechanical antenna arrays with a computational deep learning system. Instead of physically implementing 2-D or 3-D antenna structures that would increase device complexity, the system uses a 1-D array combined with neural network processing to achieve the same elevation resolution capability through software rather than hardware complexity.
Solution Approach 2:
The system creates a computational representation (copy) of 3-D spatial information from 1-D radar data through the trained neural network. The deep learning model learns to map simple radar returns to complex depth maps that encode elevation, azimuth, and range information, effectively copying 3-D spatial understanding without requiring 3-D hardware.
3Measurement precision
If deep learning models are trained with compensated ground truth depth maps from 2D range sensors, then the measurement precision is improved, but the loss of information increases due to the need for additional sensor data
Solution Approach 1:
The compensated ground truth depth maps from 2D range sensors serve as an intermediary for training the neural network. These pre-processed target depth maps, which have been corrected for ground depth, provide the training data needed for the model to learn accurate elevation prediction, enabling the system to achieve high measurement precision without requiring complex real-time sensor fusion during operation.
Data Source
AI summary
Systems and methods for resolving elevation ambiguity include acquiring, using a 1-D horizontal radar antenna array, a radar frame with range and azimuth information, and predicting a target elevation based on the frame by computing a depth map with a plurality of target depths assigned to corresponding azimuth-elevation pairs. Computing the depth map includes processing the radar frame with an encoder-decoder structured deep convolutional neural network (CNN). The CNN may be trained with a dataset including training radar frames acquired in a number of environments, and compensated ground truth depth maps associated with those environments. The compensated ground truth depth maps may be generated by subtracting a ground-depth from a corresponding ground truth depth map. The ground truth depth maps may be acquired with a 2-D range sensor, such as a LiDAR sensor, a 2-D radar sensor, and/or an IR sensor. The radar frame may also include Doppler data.