3D Radar Neural Network for Moving and Stationary Obstacle Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional perception systems for autonomous vehicles rely heavily on cameras and LIDAR sensors, which are unreliable in occluded scenes and inclement weather, and RADAR sensors struggle to distinguish between obstacles and background noise, especially in cluttered environments.
Innovation Solution
A deep neural network (DNN) using RADAR data is trained to detect moving and stationary obstacles by accumulating, ego-motion-compensating, and orthographically projecting RADAR detections, with a common trunk and separate heads for class confidence and instance regression, and uses LIDAR data for ground truth training to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional RADAR detection techniques are used to detect moving objects, then moving reflective objects can be detected, but stationary objects cannot be distinguished from background noise
Solution Approach 1:
The detection task is segmented into two separate heads: a class confidence head for detecting moving objects and an instance regression head for detecting stationary objects. This segmentation allows each head to specialize in one type of detection, improving overall accuracy for both moving and stationary object classification
Solution Approach 2:
The system transitions from traditional 2D RADAR imaging to 3D spatial representation by incorporating depth information and spatial coordinates. This dimensional enhancement allows the neural network to better distinguish stationary objects from background noise by analyzing their spatial distribution patterns
2Reliability
If conventional perception systems use cameras or LIDAR sensors, then obstacle detection can be performed, but the system becomes unreliable in occluded scenes and inclement weather
Solution Approach 1:
The system replaces expensive and weather-sensitive LIDAR and camera sensors with more robust RADAR sensors that are less affected by occlusions and inclement weather. While RADAR has its own limitations, it provides more reliable detection in adverse conditions without requiring complex multi-sensor fusion systems
3Measurement precision
If conventional detection techniques are used to extract accurate 3D information, then obstacle detection can be achieved, but run-time increases and computational demands increase
Solution Approach 1:
The system performs preliminary processing of RADAR data by accumulating detections over time and applying ego-motion compensation before feeding data to the neural network. This preliminary action reduces the computational burden during real-time inference while maintaining accurate 3D object detection
Solution Approach 2:
The system replaces complex post-processing algorithms with a trained neural network that directly outputs 3D object information. This substitution of mechanical processing with learned models reduces computational complexity and improves processing efficiency while maintaining detection accuracy
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
The DNN effectively distinguishes between obstacles and background noise, particularly in urban environments, enabling reliable obstacle detection in various weather conditions, and provides accurate dimensions and orientations of obstacles for safe autonomous vehicle navigation.
Implementation Method 1
RADAR sensors may be used to detect moving, reflective objects
Implementation Method 2
LIDAR sensors to detect obstacles in a scene
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 both highway and urban scenarios. RADAR detections may be accumulated, ego-motion-compensated, orthographically projected, and fed into a neural network(s). The neural network(s) may include a common trunk with a feature extractor and several heads that predict different outputs such as a class confidence head that predicts a confidence map and an instance regression head that predicts object instance data for detected objects. The outputs may be decoded, filtered, and/or clustered to form bounding shapes identifying the location, size, and/or orientation of detected object instances. The detected object instances may be provided to an autonomous vehicle drive stack to enable safe planning and control of the autonomous vehicle.