Radar Object Detection With Multi-Scale Feature Pyramids
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional radar systems face challenges in distinguishing between multiple reflectors, especially when dealing with stationary objects, leading to computationally expensive high-resolution techniques and difficulty in object detection due to large fields of view.
Innovation Solution
A method involving a computer-implemented approach that provides scaled data to multiple detectors, utilizing a feature pyramid of spatial scales for object detection, leveraging radar or ultrasonic signals, and employing neural networks to enhance detection efficiency and accuracy by processing range, velocity, and angular information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional Fourier transform methods are used to create range-Doppler maps with large field of view, then the system can cover a wide angular range (−75 to +75 degrees), but it becomes difficult to distinguish between multiple reflectors and requires computationally expensive high-resolution techniques
Solution Approach 1:
The patent divides the detection task into multiple detectors, each responsible for a specific angular region. This segmentation allows each detector to focus on a narrower field of view, improving reflector distinction without requiring the entire system to use computationally expensive high-resolution techniques across the full angular range.
Solution Approach 2:
The patent applies different detection strategies to different spatial regions. By assigning specialized detectors to specific angular regions, the system optimizes detection quality locally in each region while maintaining overall wide coverage, avoiding the need to apply high-complexity processing uniformly across all regions.
2Reliability
If the ego vehicle stands still and uses conventional detection methods, then stationary objects can be detected, but many points appear in the zero-th Doppler bin making it difficult to distinguish between multiple reflectors
Solution Approach 1:
The patent segments the detection space into multiple angular regions with dedicated detectors. This segmentation separates reflectors that would otherwise appear as indistinguishable points in the zero-th Doppler bin, enabling reliable distinction between multiple stationary reflectors even when the ego vehicle is stationary.
Solution Approach 2:
The patent introduces angular information as an intermediary parameter to distinguish between multiple reflectors. By incorporating angular coordinates alongside range and velocity information, the system can differentiate between multiple reflectors in the zero-th Doppler bin that would otherwise be indistinguishable.
3Measurement precision
If high-resolution techniques are deployed to resolve multiple reflection points, then reflector distinction improves, but computational cost increases significantly
Solution Approach 1:
The patent segments the detection task across multiple detectors with specialized angular regions. This segmentation achieves high reflector resolution in each local region without requiring computationally expensive high-resolution techniques to be applied globally, thereby maintaining processing efficiency while improving measurement precision.
Solution Approach 2:
The patent applies high-resolution detection capabilities partially, only in specific angular regions where needed, rather than uniformly across the entire field of view. This partial application of high-resolution processing achieves sufficient reflector distinction while avoiding the excessive computational cost of applying high-resolution techniques everywhere.
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
Enhances object detection efficiency and accuracy by processing scaled data through a feature pyramid, allowing for the detection of objects of varying sizes and properties without manual feature design, and enabling efficient training of machine learning models.
Implementation Method 1
The signal representation data is based on at least one of radar signals or ultrasonic signals. For example, the signal representation data may be based on information (for example energy or frequency) of a received reflection of a transmitted signal.
Implementation Method 2
the amount of reflected energy is highly dependent on the objects material, orientation, geometrical shape and size
Implementation Method 3
Based on a phase shift that is derivable from the received signal and its frequency, a velocity of the object may be determined (for example based on a Doppler shift in the frequency of the reflected signal).
Implementation Method 4
By determining a difference in time of arrival between a reflected signal received at various antennas, the direction of arrival (in other words: angle) may be determined.
Data Source
AI summary
A computer implemented method for detecting objects includes providing signal representation data comprising range information, velocity information and angular information; for each of a plurality of spatial scales, determining respective scaled data for the respective spatial scale based on the signal representation data, to obtain a plurality of scaled data; providing the plurality of scaled data to a plurality of detectors; and each detector carrying out object detection based on at least one of the plurality of scaled data.