Vehicle Point Cloud Target Detection With Sparsity-Adaptive Convolution

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional target detection methods are inadequate for accurately and timely detecting remote targets, especially at high speeds, leading to insufficient safety distances and increased driving risks.

Innovation Solution

A target detection method that utilizes sparse convolutions and attention mechanisms to determine convolution dilation rates based on point cloud grid sparsity, extracting and fusing global features from both remote and close targets, enhancing detection accuracy and reducing calculation time.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional target detection methods are used, then close targets can be detected, but remote targets cannot be detected accurately and timely

Engineering Contradiction:
Improvetarget detection accuracyVSAvoiddetection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent divides the point cloud space into multiple three-dimensional point cloud grids, allowing different regions (remote and close targets) to be processed independently. This segmentation enables the system to handle remote targets without compromising the detection of close targets, resolving the contradiction between detection accuracy and time loss.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a hierarchical structure with super grids at different resolution levels. By transforming point cloud grids into super grids and processing them at multiple dimensions, the system can detect remote targets accurately while maintaining real-time performance, addressing both accuracy and time requirements.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If higher detection accuracy for remote targets is achieved, then safety distance can be ensured, but computational load increases

Engineering Contradiction:
Improveremote target detection accuracyVSAvoidcomputational load
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent extracts and processes only the essential features from point cloud grids by determining sparsity and selectively applying dilated convolutions. This extraction approach reduces computational load while maintaining detection accuracy for remote targets, resolving the contradiction between accuracy and energy consumption.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent dynamically adjusts convolution dilation rates based on the sparsity of point cloud grids. By changing the dilation rate parameter according to data distribution, the system optimizes computational resources, achieving high accuracy for remote targets without excessive computational load.

Inventive Principle:
Principle #35Parameter changes

3Length of stationary object

If convolution dilation rate is increased to detect remote targets, then detection range improves, but calculation time increases

Engineering Contradiction:
Improvedetection rangeVSAvoidcalculation time
Core Design Contradiction:
Length of stationary objectVSLoss of time

Solution Approach 1:

The patent makes the convolution dilation rate dynamic by adjusting it based on the sparsity of point cloud grids. This dynamic adjustment allows the system to extend detection range when needed while reducing calculation time when sparsity is low, resolving the contradiction between detection range and calculation time.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4283571B1Target detection method, computer device, computer-readable storage medium, and vehicle
Publication Date: 2025.09.03 ANHUI NIO AUTONOMOUS DRIVING TECH CO LTD
  • EP4283571B1 patent drawingFigure 1
  • EP4283571B1 patent drawingFigure 2
  • EP4283571B1 patent drawingFigure 3

AI summary

The disclosure relates to the technical field of autonomous driving, and specifically provides a target detection method, a computer device, a computer-readable storage medium, and a vehicle, to solve the problem of detecting a target in a timely and accurate manner. For this purpose, the method of the disclosure includes: rasterizing point cloud space of three-dimensional point clouds in a vehicle driving environment to form a plurality of three-dimensional point cloud grids, and using point cloud grids including three-dimensional point clouds as target point cloud grids; determining a convolution dilation rate based on sparsity of the target point cloud grid; dilating a sparse convolution based on the convolution dilation rate; extracting a point cloud grid feature of the target point cloud grid by using a dilated sparse convolution; weighting the point cloud grid feature by using an attention mechanism to obtain a global point cloud feature; and performing target detection based on the global point cloud feature. In this way, both a remote target and a close target can be accurately detected. In addition, using the sparse convolution for detection can reduce the calculation amount and improve the detection efficiency.