LiDAR BEV 3D Object Detection for Sparse Point Clouds

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current autonomous driving systems face challenges in accurately detecting and tracking multiple objects in complex environments using LiDAR data, particularly due to sparse panoramic information, occlusion, noise, and limited field of view, which affect the robustness and accuracy of 3D object detection.

Innovation Solution

A single-stage 3D multi-object detection apparatus using a LiDAR sensor that projects raw point cloud data into Bird's Eye View (BEV) images, encoding features like height, density, intensity, and distance, and employs a convolutional neural network (CNN) to extract object centers, offsets, and 3D bounding boxes without predefined anchors, utilizing balanced L1 loss for Z-axis and size regression.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Area of stationary object

If LiDAR sensor is used to provide panoramic sparse measurements, then field of view and detection range are improved, but measurement precision and point density are worsened due to sparse panoramic information

Engineering Contradiction:
Improvefield of viewVSAvoidpoint density
Core Design Contradiction:
Area of stationary objectVSMeasurement precision

Solution Approach 1:

The patent transforms sparse 3D LiDAR point cloud data into a 2D Bird's Eye View image representation. This dimensionality change allows the system to leverage 2D image processing techniques while maintaining 3D spatial information, effectively bridging the gap between sparse panoramic measurements and dense feature extraction needed for accurate object detection

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

Solution Approach 2:

The patent combines multiple feature types (depth, intensity, height, density) into a composite BEV image representation. By fusing these different modalities of information into a unified 2D feature space, the system compensates for the sparsity of individual LiDAR measurements and achieves more robust object detection performance

Inventive Principle:
Principle #40Composite materials

2Measurement precision

If traditional multi-stage detection methods are used, then measurement precision is improved, but processing speed and real-time performance are worsened

Engineering Contradiction:
Improvedetection accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent merges the proposal generation and object detection stages into a single unified operation. By directly predicting 3D bounding boxes from BEV images without requiring separate region proposal steps, the system achieves real-time processing speeds while maintaining detection accuracy through the power of deep learning feature extraction

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent replaces traditional mechanical multi-stage detection pipelines with a streamlined deep learning-based single-stage approach. The CNN architecture automatically learns optimal feature representations and detection parameters, eliminating the need for hand-crafted features and multiple processing stages while achieving both speed and accuracy

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Ease of operation

If anchor-based methods are used, then ease of operation is improved, but adaptability and detection accuracy are worsened due to predefined anchors

Engineering Contradiction:
Improveimplementation simplicityVSAvoiddetection accuracy
Core Design Contradiction:
Ease of operationVSAdaptability or versatility

Solution Approach 1:

The patent inverts the traditional anchor-based approach by eliminating predefined anchors entirely. Instead of searching for objects within fixed anchor boxes, the system directly regresses 3D bounding box parameters from detected object centers in the BEV image, allowing flexible adaptation to objects of any size, shape, or orientation without being constrained by anchor definitions

Inventive Principle:
Principle #13The other way round (Inversion)

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 apparatus achieves accurate and efficient 3D object detection with improved robustness and speed, suitable for real-time autonomous driving applications, by leveraging compact CNN architectures and anchor-free methods.

Implementation Method 1

detection technology using Light Detector and Ranging (LiDAR) has become increasingly popular, providing sparse panoramic information of the environment

Methodology Applied
Scientific EffectLight Detector and Ranging (LiDAR): LIDAR

Implementation Method 2

a convolutional neural network (CNN) to extract object centers, offsets, and 3D bounding boxes

Methodology Applied
Scientific EffectConvolutional Neural Network processing: Image Processing

Data Source

PatentEP4138044B1Single-stage 3-dimension multi-object detecting apparatus and method for autonomous driving
Publication Date: 2025.11.26 CHUNGBUK NAT UNIV IND ACADEMIC COOP FOUNDATION
  • EP4138044B1 patent drawingFigure 1
  • EP4138044B1 patent drawingFigure 2
  • EP4138044B1 patent drawingFigure 3(a)~3(d)

AI summary

According to at least one embodiment, the present disclosure provides an apparatus for single-stage three-dimensional, 3D, multi-object detection by using a LiDAR sensor to detect 3D multiple objects, comprising: a data input module configured to receive raw point cloud data from the LiDAR sensor; a BEV image generating module configured to generate bird's eye view, BEV, images from the raw point cloud data; a learning module configured to perform a deep learning algorithm-based learning task to extract a fine-grained feature image from the BEV images; and a localization module configured to perform a regression operation and a localization operation to find 3D candidate boxes and classes corresponding to the 3D candidate boxes for detecting 3D objects from the fine-grained feature image.