GPU Radar Data Filtering for Faster Object Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Autonomous vehicles (AVs) face challenges in processing large amounts of noisy RADAR data, which is computationally intensive and slows down processing times, both in simulation environments and real-world driving scenarios.

Innovation Solution

The use of parallel processing techniques, specifically with graphical processing units (GPUs), to filter RADAR detections by determining threshold curves and reducing data dimensions, along with importance sampling to reduce noise and data size, enhances the efficiency of RADAR data processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional sequential processing methods are used to filter RADAR data, then processing accuracy can be maintained, but processing time increases and throughput decreases

Engineering Contradiction:

ImproveRADAR data processing throughputVSAvoidProcessing time

Core Design Contradiction:

ProductivityVSLoss of time

Solution Approach 1:

The patent segments the RADAR data processing pipeline into distinct stages: data reception, threshold curve calculation, data reduction, noise filtering, and object detection. Each stage is independently optimized and can be processed in parallel, particularly using GPU acceleration for computationally intensive operations like threshold curve calculation and data reduction.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary actions by pre-calculating threshold curves and pre-reducing data dimensions before main processing. Importance sampling is applied in advance to identify and retain only the most significant data points, reducing the overall data volume that requires intensive processing and thereby decreasing total processing time.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If comprehensive RADAR data is processed without reduction, then detection accuracy is improved, but computational complexity and processing load increase

Engineering Contradiction:

ImproveObject detection accuracyVSAvoidComputational complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts and removes redundant information from the RADAR data through importance sampling and dimensionality reduction techniques. By identifying and retaining only the most significant data points that contribute to object detection accuracy, the system reduces computational complexity while maintaining detection performance.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes parameters by dynamically adjusting threshold curves based on environmental conditions and data characteristics. These parameter changes enable adaptive filtering that maintains detection accuracy across varying scenarios while optimizing processing efficiency by adjusting the stringency of filtering based on current operational context.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12221128B2Reducing and processing simulated and real-world radar data

Publication Date: 2025.02.11 GM CRUISE HOLDINGS LLC

AI summary

A system for filtering RADAR data includes one or more graphical processing units (GPUs) for performing steps of a filtering process in parallel. For example, a first GPU or first portion of GPU circuitry calculates a threshold parabola for a tensor of RADAR data. In parallel, a second GPU or second portion of GPU circuitry separately reduces and indexes the tensor for comparison to the threshold parabola. The threshold parabola is compared to reduced and indexed data to filter the RADAR data. Importance sampling can also be used to reduce data, e.g., if the RADAR data includes four dimensions (range, Doppler, azimuth, and elevation).