RADAR Obstacle Detection via Orthographic Projection and Deep Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional autonomous vehicle perception systems relying on cameras and LIDAR sensors face challenges in heavy occlusions, inclement weather, and high computational demands, and RADAR systems struggle to distinguish obstacles from background noise in cluttered environments.
Innovation Solution
A deep neural network (DNN) system using RADAR data for object detection, which accumulates and ego-motion-compensates RADAR detections, projects them orthographically, and feeds them into a neural network with a common trunk and multiple heads to predict object presence, location, dimensions, and orientation, enabling the differentiation of obstacles from background noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional camera or LIDAR sensors are used for obstacle detection, then detection capability in clear conditions is achieved, but the system becomes unreliable in heavy occlusions and inclement weather
Solution Approach 1:
The patent replaces camera and LIDAR sensors with RADAR sensors for obstacle detection. RADAR uses electromagnetic wave reflection principles to detect objects, making it immune to occlusions and weather conditions that affect optical and laser-based systems. The RADAR sensor captures reflected radio waves from targets, providing reliable detection in fog, rain, snow, and heavy occlusion scenarios where cameras and LIDAR fail.
2Reliability
If conventional RADAR detection techniques are used, then moving reflective objects can be detected, but the system fails to distinguish stationary objects from background noise in cluttered environments
Solution Approach 1:
The patent transforms 3D RADAR point cloud data into 2D orthographic projection images, adding a dimensional transformation that enables application of computer vision techniques. This dimensional change allows the system to process RADAR data using convolutional neural networks originally designed for 2D image analysis, improving the ability to distinguish obstacles from background clutter through spatial pattern recognition in the projected view.
Solution Approach 2:
The patent introduces an intermediary processing stage that converts 3D RADAR detections into 2D orthographic projections before feeding them to the neural network. This intermediary representation serves as a bridge between raw RADAR data and the neural network's processing capabilities, enabling effective discrimination of obstacles from background noise through the projected spatial relationships.
3Measurement precision
If conventional detection techniques with heavy post-processing are used, then accurate 3D information can be extracted, but the run-time increases and computational efficiency decreases
Solution Approach 1:
The patent replaces conventional post-processing algorithms with a deep neural network that performs obstacle detection and 3D information extraction in an integrated manner. The neural network processes the orthographic projection images and directly outputs detection results with 3D information, eliminating the need for separate, computationally intensive post-processing steps while maintaining or improving accuracy.
Solution Approach 2:
The patent changes the processing approach by transforming the input data format (to orthographic projections) and using neural network parameters to learn optimal detection strategies. This parameter-based approach allows the system to achieve accurate 3D information extraction with reduced computational overhead compared to conventional algorithmic post-processing methods.
4Speed
If conventional RADAR techniques are used, then moving objects can be detected, but the accuracy in predicting object classification, dimension, and orientation is limited
Solution Approach 1:
The patent introduces orthographic projection as an intermediary representation that preserves spatial relationships and object characteristics in a 2D format. This intermediate representation enables the neural network to more accurately predict object classification, dimensions, and orientation by learning from the projected spatial patterns, improving upon conventional RADAR techniques that directly process 3D point clouds.
Solution Approach 2:
The patent replaces conventional RADAR processing algorithms with a deep neural network that learns optimal features for object classification and parameter estimation. The neural network processes the orthographic projection images and simultaneously predicts multiple object attributes (classification, dimension, orientation) with higher accuracy than traditional algorithmic approaches.
Data Source
AI summary
In various examples, a deep neural network(s) (e.g., a convolutional neural network) may be trained to detect moving and stationary obstacles from RADAR data of a three dimensional (3D) space. In some embodiments, ground truth training data for the neural network(s) may be generated from LIDAR data. More specifically, a scene may be observed with RADAR and LIDAR sensors to collect RADAR data and LIDAR data for a particular time slice. The RADAR data may be used for input training data, and the LIDAR data associated with the same or closest time slice as the RADAR data may be annotated with ground truth labels identifying objects to be detected. The LIDAR labels may be propagated to the RADAR data, and LIDAR labels containing less than some threshold number of RADAR detections may be omitted. The (remaining) LIDAR labels may be used to generate ground truth data.